(Holm T., Jorgensen P., Rouquier R. (Eds.) ) Triang (BookFi)

This page intentionally left blank
LONDON MATHEMATICAL SOCIETY LECTURE NOTE SERIES

Managing Editor: Professor M. Reid, Mathematics Institute,
University of Warwick, Coventry CV4 7AL, United Kingdom
The titles below are available from booksellers, or from Cambridge University Press at www.cambridge.org/mathematics
224 Computability, enumerability, unsolvability, S. B. COOPER, T. A. SLAMAN & S. S. WAINER (eds)

225 A mathematical introduction to string theory, S. ALBEVERIO et al.
226 Novikov conjectures, index theorems and rigidity I, S.C. FERRY, A. RANICKI & J. ROSENBERG (eds)
227 Novikov conjectures, index theorems and rigidity II, S.C. FERRY, A. RANICKI & J. ROSENBERG (eds)
228 Ergodic theory of Z d actions, M. POLLICOTT & K. SCHMIDT (eds)
229 Ergodicity for infinite dimensional systems, G. DA PRATO & J. ZABCZYK
230 Prolegomena to a middlebrow arithmetic of curves of genus 2, J. W. S. CASSELS & E. V. FLYNN
231 Semigroup theory and its applications, K. H. HOFMANN & M. W. MISLOVE (eds)
232 The descriptive set theory of Polish group actions, H. BECKER & A. S. KECHRIS
233 Finite fields and applications, S. COHEN & H. NIEDERREITER (eds)
234 Introduction to subfactors, V. JONES & V. S. SUNDER
235 Number theory: Séminaire de théorie des nombres de Paris 1993–94, S. DAVID (ed.)
236 The James forest, H. FETTER & B. G. DE BUEN
237 Sieve methods, exponential sums, and their applications in number theory, G. R. H. GREAVES et al.
238 Representation theory and algebraic geometry, A. MARTSINKOVSKY & G. TODOROV (eds)
240 Stable groups, F. O. WAGNER
241 Surveys in combinatorics, 1997, R. A. BAILEY (ed.)
242 Geometric Galois actions I, L. SCHNEPS & P. LOCHAK (eds)
243 Geometric Galois actions II, L. SCHNEPS & P. LOCHAK (eds)
244 Model theory of groups and automorphism groups, D. M. EVANS (ed.)
245 Geometry, combinatorial designs and related structures, J. W. P. HIRSCHFELD et al.
246 p-Automorphisms of finite p-groups, E. I. KHUKHRO
247 Analytic number theory, Y. MOTOHASHI (ed.)
248 Tame topology and O-minimal structures, L. VAN DEN DRIES
249 The atlas of finite groups: Ten years on, R. CURTIS & R. WILSON (eds)
250 Characters and blocks of finite groups, G. NAVARRO
251 Gröbner bases and applications, B. BUCHBERGER & F. WINKLER (eds)
252 Geometry and cohomology in group theory, P. KROPHOLLER, G. NIBLO, R. STÖHR (eds)
253 The q-Schur algebra, S. DONKIN
254 Galois representations in arithmetic algebraic geometry, A. J. SCHOLL & R. L. TAYLOR (eds)
255 Symmetries and integrability of difference equations, P. A. CLARKSON & F. W. NIJHOFF (eds)
256 Aspects of Galois theory, H. VÖLKLEIN et al.
257 An introduction to noncommutative differential geometry and its physical applications (2nd edition), J. MADORE
258 Sets and proofs, S. B. COOPER & J. TRUSS (eds)
259 Models and computability, S. B. COOPER & J. TRUSS (eds)
260 Groups St Andrews 1997 in Bath, I, C. M. CAMPBELL et al.
261 Groups St Andrews 1997 in Bath, II, C. M. CAMPBELL et al.
262 Analysis and logic, C. W. HENSON, J. IOVINO, A. S. KECHRIS & E. ODELL
263 Singularity theory, B. BRUCE & D. MOND (eds)
264 New trends in algebraic geometry, K. HULEK, F. CATANESE, C. PETERS & M. REID (eds)
265 Elliptic curves in cryptography, I. BLAKE, G. SEROUSSI & N. SMART
267 Surveys in combinatorics, 1999, J. D. LAMB & D. A. PREECE (eds)
268 Spectral asymptotics in the semi-classical limit, M. DIMASSI & J. SJÖSTRAND
269 Ergodic theory and topological dynamics, M. B. BEKKA & M. MAYER
271 Singular perturbations of differential operators, S. ALBEVERIO & P. KURASOV
272 Character theory for the odd order theorem, T. PETERFALVI. Translated by R. SANDLING
273 Spectral theory and geometry, E. B. DAVIES & Y. SAFAROV (eds)
274 The Mandelbrot set, theme and variations, TAN LEI (ed.)
275 Descriptive set theory and dynamical systems, M. FOREMAN et al.
276 Singularities of plane curves, E. CASAS-ALVERO
277 Computational and geometric aspects of modern algebra, M. D. ATKINSON et al.
278 Global attractors in abstract parabolic problems, J. W. CHOLEWA & T. DLOTKO
279 Topics in symbolic dynamics and applications, F. BLANCHARD, A. MAASS & A. NOGUEIRA (eds)
280 Characters and automorphism groups of compact Riemann surfaces, T. BREUER
281 Explicit birational geometry of 3-folds, A. CORTI & M. REID (eds)
282 Auslander–Buchweitz approximations of equivariant modules, M. HASHIMOTO
283 Nonlinear elasticity, Y. FU & R. OGDEN (eds)
284 Foundations of computational mathematics, R. DEVORE, A. ISERLES & E. SÜLI (eds)
285 Rational points on curves over finite fields, H. NIEDERREITER & C. XING
286 Clifford algebras and spinors (2nd edition), P. LOUNESTO
287 Topics on Riemann surfaces and Fuchsian groups, E. BUJALANCE et al.
288 Surveys in combinatorics, 2001, J. HIRSCHFELD (ed.)
289 Aspects of Sobolev-type inequalities, L. SALOFF-COSTE
290 Quantum groups and Lie theory, A. PRESSLEY (ed)
291 Tits buildings and the model theory of groups, K. TENT (ed)
292 A quantum groups primer, S. MAJID
293 Second order partial differential equations in Hilbert spaces, G. DA PRATO & J. ZABCZYK
294 Introduction to operator space theory, G. PISIER
295 Geometry and integrability, L. MASON & Y. NUTKU (eds)
296 Lectures on invariant theory, I. DOLGACHEV
297 The homotopy category of simply connected 4-manifolds, H.-J. BAUES
298 Higher operads, higher categories, T. LEINSTER (ed)
299 Kleinian groups and hyperbolic 3-manifolds, Y. KOMORI, V. MARKOVIC & C. SERIES (eds)
300 Introduction to Möbius differential geometry, U. HERTRICH-JEROMIN
301 Stable modules and the D(2)-problem, F.E.A. JOHNSON
302 Discrete and continuous nonlinear Schrödinger systems, M.J. ABLOWITZ, B. PRINARI & A.D. TRUBATCH
303 Number theory and algebraic geometry, M. REID & A. SKOROBOGATOV (eds)
304 Groups St Andrews 2001 in Oxford I, C.M. CAMPBELL, E.F. ROBERTSON & G.C. SMITH (eds)
305 Groups St Andrews 2001 in Oxford II, C.M. CAMPBELL, E.F. ROBERTSON & G.C. SMITH (eds)
306 Geometric mechanics and symmetry, J. MONTALDI & T. RATIU (eds)
307 Surveys in combinatorics 2003, C.D. WENSLEY (ed.)
308 Topology, geometry and quantum field theory, U.L. TILLMANN (ed)
309 Corings and comodules, T. BRZEZINSKI & R. WISBAUER
310 Topics in dynamics and ergodic theory, S. BEZUGLYI & S. KOLYADA (eds)
311 Groups: topological, combinatorial and arithmetic aspects, T.W. MÜLLER (ed)
312 Foundations of computational mathematics, Minneapolis 2002, F. CUCKER et al (eds)
313 Transcendental aspects of algebraic cycles, S. MÜLLER-STACH & C. PETERS (eds)
314 Spectral generalizations of line graphs, D. CVETKOVIĆ, P. ROWLINSON & S. SIMIĆ
315 Structured ring spectra, A. BAKER & B. RICHTER (eds)
316 Linear logic in computer science, T. EHRHARD, P. RUET, J.-Y. GIRARD & P. SCOTT (eds)
317 Advances in elliptic curve cryptography, I.F. BLAKE, G. SEROUSSI & N.P. SMART (eds)
318 Perturbation of the boundary in boundary-value problems of partial differential equations, D. HENRY
319 Double affine Hecke algebras, I. CHEREDNIK
320 L-functions and Galois representations, D. BURNS, K. BUZZARD & J. NEKOVÁŘ (eds)
321 Surveys in modern mathematics, V. PRASOLOV & Y. ILYASHENKO (eds)
322 Recent perspectives in random matrix theory and number theory, F. MEZZADRI & N.C. SNAITH (eds)
323 Poisson geometry, deformation quantisation and group representations, S. GUTT et al (eds)
324 Singularities and computer algebra, C. LOSSEN & G. PFISTER (eds)
325 Lectures on the Ricci flow, P. TOPPING
326 Modular representations of finite groups of Lie type, J.E. HUMPHREYS
327 Surveys in combinatorics 2005, B.S. WEBB (ed)
328 Fundamentals of hyperbolic manifolds, R. CANARY, D. EPSTEIN & A. MARDEN (eds)
329 Spaces of Kleinian groups, Y. MINSKY, M. SAKUMA & C. SERIES (eds)
330 Noncommutative localization in algebra and topology, A. RANICKI (ed)
331 Foundations of computational mathematics, Santander 2005, L.M. PARDO, A. PINKUS, E. SÜLI &
M.J. TODD (eds)
332 Handbook of tilting theory, L. ANGELERI HÜGEL, D. HAPPEL & H. KRAUSE (eds)
333 Synthetic differential geometry (2nd Edition), A. KOCK
334 The Navier–Stokes equations, N. RILEY & P. DRAZIN
335 Lectures on the combinatorics of free probability, A. NICA & R. SPEICHER
336 Integral closure of ideals, rings, and modules, I. SWANSON & C. HUNEKE
337 Methods in Banach space theory, J.M.F. CASTILLO & W.B. JOHNSON (eds)
338 Surveys in geometry and number theory, N. YOUNG (ed)
339 Groups St Andrews 2005 I, C.M. CAMPBELL, M.R. QUICK, E.F. ROBERTSON & G.C. SMITH (eds)
340 Groups St Andrews 2005 II, C.M. CAMPBELL, M.R. QUICK, E.F. ROBERTSON & G.C. SMITH (eds)
341 Ranks of elliptic curves and random matrix theory, J.B. CONREY, D.W. FARMER, F. MEZZADRI &
N.C. SNAITH (eds)
342 Elliptic cohomology, H.R. MILLER & D.C. RAVENEL (eds)
343 Algebraic cycles and motives I, J. NAGEL & C. PETERS (eds)
344 Algebraic cycles and motives II, J. NAGEL & C. PETERS (eds)
345 Algebraic and analytic geometry, A. NEEMAN
346 Surveys in combinatorics 2007, A. HILTON & J. TALBOT (eds)
347 Surveys in contemporary mathematics, N. YOUNG & Y. CHOI (eds)
348 Transcendental dynamics and complex analysis, P.J. RIPPON & G.M. STALLARD (eds)
349 Model theory with applications to algebra and analysis I, Z. CHATZIDAKIS, D. MACPHERSON, A. PILLAY &
A. WILKIE (eds)
350 Model theory with applications to algebra and analysis II, Z. CHATZIDAKIS, D. MACPHERSON, A. PILLAY &
A. WILKIE (eds)
351 Finite von Neumann algebras and masas, A.M. SINCLAIR & R.R. SMITH
352 Number theory and polynomials, J. MCKEE & C. SMYTH (eds)
353 Trends in stochastic analysis, J. BLATH, P. MÖRTERS & M. SCHEUTZOW (eds)
354 Groups and analysis, K. TENT (ed)
355 Non-equilibrium statistical mechanics and turbulence, J. CARDY, G. FALKOVICH & K. GAWEDZKI
356 Elliptic curves and big Galois representations, D. DELBOURGO
357 Algebraic theory of differential equations, M.A.H. MACCALLUM & A.V. MIKHAILOV (eds)
358 Geometric and cohomological methods in group theory, M.R. BRIDSON, P.H. KROPHOLLER & I.J. LEARY (eds)
359 Moduli spaces and vector bundles, L. BRAMBILA-PAZ, S.B. BRADLOW, O. GARCÍA-PRADA &
S. RAMANAN (eds)
360 Zariski geometries, B. ZILBER
361 Words: Notes on verbal width in groups, D. SEGAL
362 Differential tensor algebras and their module categories, R. BAUTISTA, L. SALMERÓN & R. ZUAZUA
363 Foundations of computational mathematics, Hong Kong 2008, F. CUCKER, A. PINKUS & M.J. TODD (eds)
364 Partial differential equations and fluid mechanics, J.C. ROBINSON & J.L. RODRIGO (eds)
365 Surveys in combinatorics 2009, S. HUCZYNSKA, J.D. MITCHELL & C.M. RONEY-DOUGAL (eds)
366 Highly oscillatory problems, B. ENGQUIST, A. FOKAS, E. HAIRER & A. ISERLES (eds)
367 Random matrices: High dimensional phenomena, G. BLOWER
368 Geometry of Riemann surfaces, F.P. GARDINER, G. GONZÁLEZ-DIEZ & C. KOUROUNIOTIS (eds)
369 Epidemics and rumours in complex networks, M. DRAIEF & L. MASSOULIÉ
370 Theory of p-adic distributions, S. ALBEVERIO, A.YU. KHRENNIKOV & V.M. SHELKOVICH
371 Conformal fractals, F. PRZYTYCKI & M. URBAŃSKI
372 Moonshine: The first quarter century and beyond, J. LEPOWSKY, J. MCKAY & M.P. TUITE (eds)
373 Smoothness, regularity, and complete intersection, J. MAJADAS & A. RODICIO
374 Geometric analysis of hyperbolic differential equation: An introduction, S. ALINHAC
375 Triangulated categories, T. HOLM, P. JØRGENSEN & R. ROUQUIER (eds)
376 Permutation patterns, S. LINTON, N. RUŠKUC & V. VATTER (eds)
377 An introduction to Galois cohomology and its applications, G. BERHUY
378 Probability and mathematical genetics, N.H. BINGHAM & C.M. GOLDIE (eds)
London Mathematical Society Lecture Note Series: 375
Triangulated Categories
Edited by
THORSTEN HOLM
Leibniz Universität Hannover, Germany
PETER JØRGENSEN
University of Newcastle upon Tyne
R A P H A Ë L RO U QU I E R
University of Oxford
cambridge university press
Cambridge, New York, Melbourne, Madrid, Cape Town, Singapore,
São Paulo, Delhi, Dubai, Tokyo
Cambridge University Press
The Edinburgh Building, Cambridge CB2 8RU, UK
Published in the United States of America by Cambridge University Press, New York
www.cambridge.org
Information on this title: www.cambridge.org/9780521744317

C Cambridge University Press 2010
This publication is in copyright. Subject to statutory exception

and to the provisions of relevant collective licensing agreements,
no reproduction of any part may take place without the written
permission of Cambridge University Press.
First published 2010
Printed in the United Kingdom at the University Press, Cambridge
A catalogue record for this publication is available from the British Library
Library of Congress Cataloguing in Publication data

Triangulated categories / edited by Thorsten Holm, Peter Jørgensen, Raphaël Rouquier.
p. cm. – (London Mathematical Society lecture note series ; 375)
Includes index.
ISBN 978-0-521-74431-7 (pbk.)
1. Triangulated categories. I. Holm, Thorsten, 1965– II. Jørgensen, Peter, 1970–
III. Rouquier, Raphaël. IV. Title. V. Series.
QA169.T685 2010
512 .62 – dc22 2010012362
ISBN 978-0-521-74431-7 Paperback
Cambridge University Press has no responsibility for the persistence or

accuracy of URLs for external or third-party internet websites referred to
in this publication, and does not guarantee that any content on such
websites is, or will remain, accurate or appropriate.
Contents
Preface page vii
Triangulated categories: definitions, properties, and examples 1

thorsten holm and peter jørgensen
Cohomology over complete intersections via exterior algebras 52
luchezar l. avramov and srikanth b. iyengar
Cluster algebras, quiver representations and triangulated
categories 76
bernhard keller
Localization theory for triangulated categories 161
henning krause
Homological algebra in bivariant K-theory and other
triangulated categories. I 236
ralf meyer and ryszard nest
Derived categories and Grothendieck duality 290
amnon neeman
Derived categories and algebraic geometry 351
rapha ël rouquier
Triangulated categories for the analysts 371
pierre schapira
Algebraic versus topological triangulated categories 389
stefan schwede
v
vi Contents
Derived categories of coherent sheaves on algebraic varieties 408

yukinobu toda
Rigid dualizing complexes via differential graded algebras
(survey) 452
amnon yekutieli
Preface
This volume grew out of a Workshop on Triangulated Categories held at the

University of Leeds in August 2006. The meeting, a Satellite of the Interna-
tional Congress of Mathematicians 2006, has been generously supported by the
Leverhulme Foundation (via the network Algebras, Representations and Appli-
cations), the London Mathematical Society (Conference Grant Ref. 1438) and
the University of Leeds.
Over the past decades, triangulated categories have made their way into
many different parts of mathematics, to the extent that today, they can be
viewed as a unifying theory underlying major parts of modern mathematics.
The Leeds workshop has brought together researchers from many parts of
mathematics who all use triangulated methods but would not usually meet
at more specialized conferences, with the aim to promote cross fertilization
leading to new applications of triangulated categories.
The present book collects surveys by leading experts reflecting a broad range
of important topics covered at the workshop. However, it is not a proceedings
volume recording precisely the talks given at the conference and it does not
claim to be a comprehensive coverage of all the numerous applications of
triangulated categories throughout mathematics.
There are contributions dealing with fundamental general aspects of trian-
gulated categories as well as articles covering important applications, e.g. in
algebraic geometry, algebraic topology, commutative algebra, algebraic analy-
sis, K-theory or representation theory.
We wish to express our sincere thanks to the authors of the contributions, as
well as to the referees.
We think that the interdisciplinary spirit of the successful Leeds workshop
and the many fruitful discussions having taken place there are well reflected by
the articles and we hope that specialists and non-specialists alike will benefit
from the broad perspective on triangulated categories and their applications
provided by the surveys.
We are very grateful to the staff at Cambridge University Press for their
help, their patience and constant support in bringing this book together.
Hannover, Newcastle, and Oxford, January 2010

Thorsten Holm, Peter Jørgensen, and Raphaël Rouquier
vii
Triangulated categories:
definitions, properties, and examples
thorsten holm and peter jørgensen
Triangulated categories were introduced in the mid 1960’s by J.L. Verdier

in his thesis, reprinted in [16]. Axioms similar to Verdier’s were indepen-
dently also suggested in [2]. Having their origins in algebraic geometry and
algebraic topology, triangulated categories have by now become indispensable
in many different areas of mathematics. Although the axioms might seem a
bit opaque at first sight it turned out that very many different objects actu-
ally do carry a triangulated structure. Nowadays there are important appli-
cations of triangulated categories in areas like algebraic geometry (derived
categories of coherent sheaves, theory of motives) algebraic topology (stable
homotopy theory), commutative algebra, differential geometry (Fukaya cate-
gories), microlocal analysis or representation theory (derived and stable module
categories).
It seems that the importance of triangulated categories in modern mathe-
matics is growing even further in recent years, with many new applications
only recently found; see B. Keller’s article in this volume for one striking
example, namely the cluster categories occurring in the context of S. Fomin
and A. Zelevinsky’s cluster algebras which have been introduced only around
2000.
In this chapter we aim at setting the scene for the survey articles in
this volume by providing the relevant basic definitions, deducing some ele-
mentary general properties of triangulated categories, and providing a few
examples.
Certainly, this cannot be a comprehensive introduction to the subject. For
more details we refer to one of the well-written textbooks on triangulated
categories, e.g. [4], [5], [7], [8], [12], [17], and for further topics also to the
surveys in this volume.
This introductory chapter should be accessible for a reader with a good
background in algebra and some basic knowledge of category theory and homo-
logical algebra.
1
2 Thorsten Holm and Peter Jørgensen
1. Additive categories
In this first section we shall discuss the fundamental notion of an additive
category and provide some examples. In particular, the category of complexes
over an additive category is introduced which will play a fundamental role in
the sequel.
Definition 1.1. A category A is called an additive category if the following

conditions hold:
(A1) For every pair of objects X, Y the set of morphisms HomA (X, Y ) is an
abelian group and the composition of morphisms
HomA (Y, Z) × HomA (X, Y ) → HomA (X, Z)
is bilinear over the integers.

(A2) A contains a zero object 0 (i.e. for every object X in A each morphism
set HomA (X, 0) and HomA (0, X) has precisely one element).
(A3) For every pair of objects X, Y in A there exists a coproduct X ⊕ Y in A.
Remark 1.2.
(i) A category satisfying (A1) and (A2) is called a preadditive category.
(ii) We recall the notion of coproduct from category theory. Let C be a category
and X, Y objects in C. A coproduct of X and Y in C is an object X ⊕ Y
together with morphisms ιX : X → X ⊕ Y and ιY : Y → X ⊕ Y satisfying
the following universal property: for every object Z in C and morphisms
fX : X → Z and fY : Y → Z there is a unique morphism f : X ⊕ Y → Z
making the following diagram commutative.
Z
fX 6 @ I f
@Y
f
@
X ι - X ⊕ Y ι Y
X Y
Example 1.3.
(i) Let R be a ring and consider R as a category CR with only one object.
The unique morphism set is the underlying abelian group and composition
of morphisms is given by ring multiplication. Then CR satisfies (A1) and
(A2), thus preadditive categories can be seen as generalizations of rings.
But CR is not additive in general; in fact the coproduct of the unique object
with itself would have to be again this object together with fixed ring
elements ι1 , ι2 , and the universal property would mean that for arbitrary
Triangulated categories 3
ring elements f1 , f2 there existed a unique element f factoring them as

f1 = f ι1 and f2 = f ι2 .
(ii) Let R be a ring (associative, with unit element). Then the category R-
Mod of all R-modules is additive. Similarly, the category R-mod of
finitely generated R-modules is additive. In particular, the categories Ab
of abelian groups and VecK of vector spaces over a field K are additive.
(iii) The full subcategory of Ab of free abelian groups is additive.
(iv) For a ring R the full subcategory R-Proj of projective R-modules is
additive; similarly for R-proj, the category of finitely generated projective
R-modules.
1.1. The category of complexes

Let A be an additive category. A complex over A is a family X = (Xn , dnX )n∈Z
where Xn are objects in A and dnX : Xn → Xn−1 are morphisms such that
dn ◦ dn+1 = 0 for all n ∈ Z. Usually, a complex is written as a sequence of
objects and morphisms as follows.
dn+1 dn
. . . → Xn+1 −→ Xn −→ Xn−1 −→ . . .
Let X = (Xn , dnX ) and Y = (Yn , dnY ) be complexes over A. A morphism of com-
plexes f : X → Y is a family of morphisms f = (fn : Xn → Yn )n∈Z satisfy-
ing dnY ◦ fn = fn−1 ◦ dnX for all n ∈ Z, i.e. we have the following commutative
diagram.
... -X -X -X - ...
n+1 n n−1
fn+1 fn fn−1
? ? ?
... -Y -Y -Y - ...
n+1 n n−1
The complexes over an additive category A together with the morphisms of

complexes form a category C(A), the category of complexes over A.
Proposition 1.4. Let A be an additive category. Then the category of complexes

C(A) is again additive.
Proof. (A1) Addition of morphisms is defined degreewise, i.e. for two mor-
phisms f = (fn )n∈Z and g = (gn )n∈Z from X to Y their sum is f + g :=
(fn + gn )n∈Z . Using the additive structure of A it is then easy to check that
(A1) holds.
(A2) The zero object in C(A) is the complex (0A , d) where 0A is the zero
object of the additive category A and all differentials are the unique (zero)
morphism on the zero object.
(A3) The coproduct of two complexes X = (Xn , dnX ) and Y = (Yn , dnY ) is
defined degreewise by using the coproduct in the additive category A. More
precisely X ⊕ Y = (Xn ⊕ Yn , dn )n∈Z where the differential is obtained by the
universal property as in the following diagram.
Xn−1 ⊕ Yn−1
ιXn−1 dnX 6 @ I ιY d Y
dn @ n−1 n
@
-
Xn ιXn Xn ⊕ Yn ιYn Yn
From uniqueness in the universal property applied to
Xn−2 ⊕ Yn−2
6 I
@
@ 0
0
dn−1 dn @
@
Xn - Xn ⊕ Yn Yn
ιXn ιYn
it follows that dn−1 ◦ dn = 0. This complex indeed satisfies the properties of a
coproduct in the category of complexes C(A), with morphisms of complexes
ιX = (ιXn )n∈Z : X → X ⊕ Y and ιY = (ιYn )n∈Z : Y → X ⊕ Y . For checking
the universal property let Z be an arbitrary complex and let fX : X → Z
and fY : Y → Z be arbitrary morphisms. The unique morphism of complexes
satisfying fX = f ◦ ιX and fY = f ◦ ιY is f = (fn )n∈Z : X ⊕ Y → Z, where
fn is obtained from the universal property in degree n as in the following
diagram.
Zn
(fX )n f 6 I
@ (fY )n
@
n
@
Xn ι - Xn ⊕ Yn ι Yn
Xn Yn
Remark 1.5. For complexes over A = R-Mod where R is a ring with unit
(and other similar examples) the coproduct of two complexes is more easily
be described on elements as X ⊕ Y = (Xn ⊕ Yn , dn )n∈Z where the differential
is given by dn (xn , yn ) = (dnX (xn ), dnY (yn )) for xn ∈ Xn and yn ∈ Yn , and with
morphisms ιX : X → X ⊕ Y and ιY : Y → X ⊕ Y being the inclusion maps.
The unique morphism of complexes satisfying fX = f ◦ ιX and fY = f ◦ ιY
is then given by fn (xn , yn ) = fX (xn ) + fY (yn ).
1.2. The homotopy category of complexes

Let A be an additive category. Morphisms f, g : X → Y in the category C(A)
of complexes are called homotopic, denoted f ∼ g, if there exists a family
(sn )n∈Z of morphisms sn : Xn → Yn+1 in A, satisfying fn − gn = dn+1 Y
sn +
sn−1 dn for all n ∈ Z.
X
In particular, setting g to be the zero morphism, we can speak of morphisms

being homotopic to zero.
It is easy to check that ∼ is an equivalence relation. Moreover, if f ∼ g :
X → Y are homotopic and α : W → X is an arbitrary morphism of complexes,
then also the compositions f α ∼ gα are homotopic. In fact, (sn αn )n∈Z are
homotopy maps since
(fn − gn )αn = (dn+1

Y
sn + sn−1 dnX )αn = dn+1
Y
(sn αn ) + (sn−1 αn−1 )dnW .
Similarly, if f, g : X → Y are homotopic and β : Y → Z is a morphism of

complexes then βf ∼ βg are homotopic.
This implies that we have a well-defined composition of equivalence classes
of morphisms modulo homotopy by defining the composition on representa-
tives.
Definition 1.6. Let A be an additive category. The homotopy category K(A)

has the same objects as the category C(A) of complexes over A. The morphisms
in the homotopy category are the equivalence classes of morphisms in C(A)
modulo homotopy, i.e.
HomK(A) (X, Y ) := HomC(A) (X, Y )/ ∼ .
Proposition 1.7. Let A be an additive category. Then the homotopy category

K(A) is again an additive category.
Proof. Addition of morphisms in K(A) is defined via addition on representa-

tives (it is an easy observation that this is well-defined) and then the sets of
morphisms HomK(A) (X, Y ) inherit the structure of an abelian group from the
category C(A) of complexes, and also bilinearity of composition. Moreover,
the zero object is the same as in C(A).
It remains to be checked that the universal property of the coproduct X ⊕ Y
in C(A) (cf. Proposition 1.4) also carries over to the homotopy category. In fact,
the equivalence classes of the morphisms ιX , ιY and f still make the relevant
diagram (cf. Remark 1.2) commutative; for uniqueness we observe that if
there is another morphism g making the diagram for the universal property
commutative in K(A), i.e. up to homotopy, then this gives a homotopy between
f and g.
2. Abelian categories
In this section we shall review the fundamental definition of an abelian category,
including the necessary background on the categorical notions of kernels and
cokernels. The prototype example of an abelian category will be the category
R-Mod of modules over a ring R; but we will also see other examples in due
course.
We first recall some notions from category theory. Let A be an additive
category; in particular for every pair of objects X, Y there is a zero morphism,
namely the composition of the unique morphisms X → 0 → Y involving the
zero object of A.
The kernel of a morphism f : X → Y is an object K together with a mor-
phism k : K → X such that
(i) f ◦ k = 0,
(ii) (universal property) for every morphism k : K → X such that f ◦ k = 0,
there is a unique morphism g : K → K making the following diagram
commutative.
Y
A K
f 6
0 A
A0
K A
k
X
YH
H A
H Ig A
@
H
k H @
HH @A
K
By the usual universal property argument, the kernel, if it exists, is unique up

to isomorphism; notation: ker f .
Dually, the cokernel of a morphism f : X → Y is an object C together with
a morphism c : Y → C such that
(i) c ◦ f = 0,
(ii) (universal property) for every morphism c : Y → C such that c ◦ f = 0,
there is a unique morphism g : C → C making the following diagram
commutative.
Y
A
f
c A
? Ac
C gA
0-
X
HH A
HH @@ A
0 H UA
HH R
@
j C
Again, the cokernel, if it exists, is unique up to isomorphism; notation:

coker f .
If the above morphism k : ker f → X has a cokernel in A, this is called the
coimage of f , and it is denoted by coim f .
If the above morphism c : Y → coker f has a kernel in A, this is called the
image of f and it is denoted by im f .
Example 2.1. Let R be a ring. In the category R-Mod of all R-modules the
categorical kernels and cokernels are the usual ones, i.e., for a morphism f :
X → Y we have ker f = {x ∈ X | f (x) = 0} and coker f = Y / im f where
im f = {f (x) | x ∈ X} is the usual image of f .
Remark 2.2. Suppose that for a morphism f both the coimage and the image
exist. Then we claim that it follows from the universal properties that there is
a natural morphism coim f → im f .
In fact, the image of f is the kernel of c : Y → coker f , hence there is
a morphism k̃ : im f → Y such that c ◦ k̃ = 0 and by the universal property
there exists a unique morphism g̃ : X → im f making the following diagram
commutative.
coker f
AK
c 6
0 A
A0
k̃ im f A
Y
YH
H
H @ I A
g̃ A
f H@ A
H
H @X
H
Note that k̃ ◦ g̃ ◦ k = f ◦ k = 0, which implies that g̃ ◦ k : ker f → im f must

be zero, by using the uniqueness in the following diagram.
coker f
AK
c 6
0 A
A0
k̃ im f A
Y H
YH
HH @ I A
A
0 H@ A
HH @
ker f
Then we can consider the following diagram for the universal property of the
coimage
X
A
k
c̃ A
? A g̃
ker f 0- coim fA
H
HH @ A
A
0 HH @ AU
HH R
@
j im f
and deduce that there is a unique morphism coim f → im f , as desired.
Definition 2.3. An additive category A is called an abelian category if the

following axioms are satisfied:
(A4) Every morphism in A has a kernel and a cokernel.

(A5) For every morphism f : X → Y in A, the natural morphism coim f →
im f is an isomorphism.
Example 2.4.
(i) Let R be a ring. The category R-Mod of all R-modules is an abelian
category. In fact, (A5) follows directly from the isomorphism theorem for
R-modules.
However, the subcategory R-mod of finitely generated modules is not
abelian in general since kernels of homomorphisms between finitely gen-
erated modules need not be finitely generated. Indeed we have that R-mod
is an abelian category if and only if R is Noetherian.
In particular, the category of finite-dimensional vector spaces over a
field is abelian, and the category of finitely generated abelian groups is
abelian.
(ii) The subcategory of Ab consisting of free abelian groups is not abelian.
On the other hand, for a prime number p, the abelian p-groups form
an abelian subcategory of Ab (an abelian group is called a p-group if for
every element a we have pk a = 0 for some k).
(iii) For finding examples of additive categories satisfying (A4) but failing to
be abelian, the following observation can be useful. Suppose f : X → Y
is a morphism with ker f = 0 and coker f = 0, i.e. a monomorphism and
an epimorphism. Then the coimage of f is the identity on X, the image
of f is the identity on Y and hence the natural morphism coim f →
im f is just f itself. So in this special case the axiom (A5) states that
a morphism which is a monomorphism and an epimorphism must be
invertible.
(iv) Explicit examples of additive categories where axiom (A5) fails for the
above reason are the category of topological abelian groups (with contin-
uous group homomorphisms) or the category of Banach complex vector
spaces (with continuous linear maps). In such categories the cokernel of
a morphism f : X → Y is of the form Y / imf where imf is the closure
of the usual set-theoretic image of f . In particular, the natural morphism
coim f → im f is the inclusion of the usual image of f into its closure,
and this is in general not an isomorphism.
Proposition 2.5. Let A be an abelian category. Then the category of complexes

C(A) is also abelian.
Proof. We have seen in Proposition 1.4 that C(A) is an additive category, so it

remains to verify the axioms (A4) and (A5).
(A4) Let f : X → Y be a morphism in C(A), i.e. f = (fn )n∈Z with fn :
Xn → Yn morphisms in A. We show the existence of a kernel and leave the
details of the dual argument for the cokernel as an exercise.
Since A is abelian, each morphism fn : Xn → Yn has a kernel Kn := ker fn
in A, coming with a morphism kn : Kn → Xn satisfying the above universal
property. Note that for every n ∈ Z we have fn−1 ◦ dnX ◦ kn = dnY ◦ fn ◦ kn = 0.
Then it follows by the universal property of kernels that there is a unique
morphism dnK : Kn → Kn−1 such that kn−1 ◦ dnK = dnX ◦ kn . Note that
kn−1 ◦ dnK ◦ dn+1

K
= dnX ◦ kn ◦ dn+1
K
= dnX ◦ dn+1
X
◦ kn+1 = 0
since X is a complex. By uniqueness of the map in the universal property of

Kn−1 it follows that dnK ◦ dn+1
K
= 0, i.e. (Kn , dnK ) is a complex.
Combining the universal properties of the kernels Kn it easily follows that
the complex (Kn , dnK ) indeed satisfies the universal property for the kernel of
f in C(A).
(A5) The crucial observation is that a morphism of complexes f = (fn ) :
X → Y is an isomorphism in C(A) if and only if each fn is an isomorphism
in A. In fact, if each fn is an isomorphism, with inverse gn , then the family
g = (gn ) is automatically a morphism of complexes (and hence clearly an
inverse to f in C(A)): for all n ∈ Z we have
X
dn+1 ◦ gn+1 = gn ◦ fn ◦ dn+1
X
◦ gn+1 = gn ◦ dn+1
Y
◦ fn+1 ◦ gn+1 = gn ◦ dn+1
Y
.
The reverse implication is obvious.

For axiom (A5) now consider the natural morphism coim f → im f . In the
proof of (A4) above we have seen that kernels and cokernels in C(A), and
hence also the morphism coim f → im f , are obtained degreewise. But since
A is abelian by assumption, we know that for every n the natural morphism
coim fn → im fn in A is indeed an isomorphism. Then, by the introductory
remark, the morphism of complexes (coim fn → im fn )n∈Z is an isomorphism
in C(A).
An important observation is that the homotopy category K(A) is not abelian

in general, even if A is abelian.
Example 2.6. We provide an explicit example for the failure of axiom (A4)
in a homotopy category. Consider the abelian category A = Ab of abelian
groups.
Let f : X → Y be the following morphism of complexes of abelian groups,
with non-zero entries in degrees 1 and 0,
... 0 - 0 - Z - 0 ...
id
? ? ? ?
... 0 - Z id- Z - 0 ...
In the category C(Ab) of complexes f is non-zero and has the zero complex
as kernel (cf. the proof of Proposition 2.5). However, f is homotopic to zero
(with the identity as homotopy map), i.e. f = 0 in the homotopy category
K(Ab).
We claim that in the homotopy category f has no kernel. Recall the cate-
gorical definition of the kernel of a morphism f : X → Y from Section 2.
Suppose for a contradiction that our morphism f had a kernel in K(Ab). So
there is a complex . . . → K1 → K0 → K−1 → . . . and a morphism k = k0 :
K0 → Z of abelian groups (in all other degrees the map k has to be zero since X
is concentrated in degree 0). The image of k, being a subgroup of Z, has the form
rZ for some fixed r ∈ Z. Now choose K = X and consider the morphisms
l : K → X given by multiplication with l for any l ∈ Z. Clearly, f ◦ l = 0 in
K(Ab) since f = 0 in K(Ab). According to the universal property of a kernel,
there must exist (unique) morphisms ul : Z → K0 such that k ◦ ul = l up to
homotopy. However, these maps are from K = X to X and this complex is
concentrated in degree 0. Thus there are no non-zero homotopy maps and so
k ◦ ul = l as morphism of abelian groups. But the image of k ◦ ul is contained
in the image of k which is rZ for a fixed r, so k ◦ ul = l can not hold for
arbitrary l ∈ Z, a contradiction.
Hence axiom (A4) fails and therefore the homotopy category K(Ab) is not
an abelian category.
3. Definition of triangulated categories

We have seen in the previous section that the homotopy category of complexes
is not abelian in general. We shall see in Section 6 below that K(A) carries the
structure of a triangulated category, a concept which we are going to define in
this section. Roughly, one should think of the distinguished triangles occurring
in this context as a replacement for short exact sequences (which do not exist
in general since K(A) is not abelian). However, for an additive category to
be abelian is purely an inherent property of the category. On the other hand
a triangulated structure is an extra piece of data, consisting of a suspension
functor and a set of distinguished triangles chosen suitably to satisfy certain
axioms. In particular, an additive category can have many different triangulated
structures; see [1] for more details and examples.
A functor between additive categories is called an additive functor if
for every pair of objects X, Y the map Hom(X, Y ) → Hom((X), (Y )) is a
homomorphism of abelian groups.
Let T be an additive category and let : T → T be an additive functor
which is an automorphism (i.e. it is invertible, thus there exists a functor −1
on T such that ◦ −1 and −1 ◦ are the identity functors).
A triangle in T is a sequence of objects and morphisms in T of the form
u v w
X −→ Y −→ Z −→ X.
A morphism of triangles is a triple (f, g, h) of morphisms such that the

following diagram is commutative in T
u- v- w-
X Y Z X
f g h f
? ? ? ?
u- v - w-
X Y Z X
If in this situation, the morphisms f, g and h are isomorphisms in T , then the

morphism of triangles is called an isomorphism of triangles.
Definition 3.1. A triangulated category is an additive category T together with

an additive automorphism , the translation or shift functor, and a collection
of distinguished triangles satisfying the following axioms
(TR0) Any triangle isomorphic to a distinguished triangle is again a distin-

guished triangle.
id
(TR1) For every object X in T , the triangle X → X → 0 → X is a distin-
guished triangle.
(TR2) For every morphism f : X → Y in T there is a distinguished triangle

f
of the form X → Y → Z → X.
u v w v
(TR3) If X → Y → Z → X is a distinguished triangle, then also Y →
w −u
Z → X → Y is a distinguished triangle, and vice versa.
u v w u v
(TR4) Given distinguished triangles X → Y → Z → X and X → Y →
w
Z → X , then each commutative diagram
u- v- w-
X Y Z X
f g f
? ? ?
u- v - w-
X Y Z X
can be completed to a morphism of triangles (but not necessarily

uniquely).
u
(TR5) (Octahedral axiom) Given distinguished triangles X → Y → Z →
v vu
X, Y → Z → X → Y and X → Z → Y → X, there exists a
distinguished triangle Z → Y → X → Z making the following
diagram commutative
u- - Z - X
X Y
idX v idX
? ? ? ?
vu- - Y - X
X Z
u idZ u
? ? ? ?
v- - X - Y
Y Z
idX
? ? ? ?
Z - Y - X - Z
Remark 3.2. The above version (TR5) of the octahedral axiom is taken from
the book by Kashiwara and Schapira [7, Sec. 1.4]. There are various other
versions appearing in the literature which are equivalent to (TR5), see for
instance A. Neeman’s article [13] or his book [12]; a short treatment can
also be found in A. Hubery’s notes [6] (which are based on the former
references).
We shall only mention two variations here. Mainly a reformulation of the
axiom (TR5) is the following. Note that in (TR5) the given three distinguished
triangles are placed in the first three rows, whereas in (TR5’) below they are
placed in the first two rows and the second column.
u v l
(TR5’) Given distinguished triangles X → Y → Z → X, Y → Z → X →
vu s
Y and X → Z → Y → X, then there exists a distinguished triangle Z →
v
Y → X → Z making the following diagram commutative and satisfying
(u)s = lv .
u- - Z - X
X Y
idX v idX
? ? ? ?
vu- - Y s - X
X Z
v
? id ?
X
X
- X
? ?
Y - Z
It is not difficult to check that (TR5) and (TR5’) are indeed equivalent; we leave
this verification as an exercise to the reader.
The following version (TR5”) of the octahedral axiom can be found in

Neeman’s book [12, Prop. 1.4.6]. It is less obvious that it is equivalent to
(TR5); for details on this we refer the reader to [12], [13] and [6].
u v
(TR5”) Given distinguished triangles X → Y → Z → X, Y → Z →
vu
X → Y and X → Z → Y → X, then there exists a distinguished tri-
angle Z → Y → X → Z making the following diagram commutative in
which every row and every column is a distinguished triangle.
u- - Z - X
X Y
idX v idX
? ? ? ?
vu- - Y - X
X Z
? ? ? ?
0 - X idX- X - 0
? ? ? ?
X u- Y - Z - 2X
4. Some formal properties of triangulated categories

We shall draw some first consequences from the definition. Let T be a triangu-
lated category with translation functor .
u v w
Proposition 4.1 (Composition of morphisms). Let X → Y → Z → X be a
distinguished triangle. Then v ◦ u = 0 and w ◦ v = 0, i.e. any composition of
two consecutive morphisms in a distinguished triangle vanishes.
Proof. By the rotation property (TR3) it suffices to show that v ◦ u = 0. Also
v w −u
by (TR3) we have a distinguished triangle Y → Z → X → Y . By (TR1)
and (TR4) the following diagram can be completed to a morphism of triangles.
v- w- −u
- Y
Y Z X
v id v
? ? ?
id- 0- 0-
Z Z 0 Z
In particular, −(v ◦ u) = −v ◦ u = 0 which implies v ◦ u = 0 since
is an automorphism.
u v w
Proposition 4.2 (Long exact sequences). Let X → Y → Z → X be a dis-
tinguished triangle. For any object T ∈ T there is a long exact sequence of
abelian groups
i u∗ i v∗
. . . → HomT (T , i X) → HomT (T , i Y ) → HomT (T , i Z)
i w∗
→ HomT (T , i+1 X) → . . .
Proof. For abbreviation we denote by f∗ := HomT (T , f ) the morphism
induced by f under the functor HomT (T , −) on the additive category T .
By the rotation property, it suffices to show that
i u∗ i v∗
HomT (T , i X) → HomT (T , i Y ) → HomT (T , i Z)
is an exact sequence of abelian groups.
By Proposition 4.1 we have i v ◦ i u = 0 and hence also i v∗ ◦ i u∗ =
0, i.e. the image of i u∗ is contained in the kernel of i v∗ .
Conversely, take f in the kernel of i v∗ . Consider the following diagram
whose rows are distinguished triangles by (TR1) and (TR3).
0- 0- − id-
−i T 0 −i+1 T −i+1 T
−i f 0 −i+1 f
? ? ?
v- w- −u-
Y Z X Y
The left hand square is commutative by assumption on f . By (TR4) there exists

a morphism h : −i+1 T → X completing the above diagram to a morphism
of triangles. In particular, −i+1 f = u ◦ h and hence f = i u ◦ i−1 h is in
the image of i u∗ as desired.
Proposition 4.3 (Triangulated 5-lemma). Suppose we are given a morphism

of distinguished triangles as in the following diagram.
u- v- w-
X Y Z X
f g h f
? ? ? ?
u- v - w-
X Y Z X
If f and g are isomorphisms then also h is an isomorphism.
Proof. We apply the functor Hom(Z , −) := HomT (Z , −) to the distinguished

triangles. By Proposition 4.2 this leads to the following commutative diagram
whose rows are exact sequences of abelian groups.
Hom(Z , X) - Hom(Z , Y ) - Hom(Z , Z) - Hom(Z , X) - Hom(Z , Y )
f∗ g∗ h∗ f∗ g∗
? ? ? ? ?
Hom(Z , X ) - Hom(Z , Y ) - Hom(Z , Z ) - Hom(Z , X ) - Hom(Z , Y )
By assumption, f and g are isomorphisms and hence also f∗ , g∗ , f∗ and

g∗ are isomorphisms. So we can appeal to the usual 5-lemma in the category
of abelian groups to deduce that h∗ is an isomorphism. In particular the identity
idZ has a preimage, i.e. there exists a morphism q ∈ HomT (Z , Z) such that
h ◦ q = idZ .
A similar argument using the functor HomT (−, Z ) produces a left inverse
to h, thus h is an isomorphism.
u v w
Proposition 4.4 (Split triangles). Let X → Y → Z → X be a distinguished
triangle where w = 0 is the zero morphism. Then the triangle splits, i.e. u is a
split monomorphism and v is a split epimorphism.
Remark 4.5. The notion of split monomorphism is synonymous with that of

a section, and a split epimorphism is also known as a retraction.
Proof. We first show that u is a split monomorphism, i.e. there exists a mor-
phism u such that u ◦ u = idX . We have the following commutative diagram
of distinguished triangles.
u- v- 0-
X Y Z X
id 0 id
? ? ?
id- 0- 0-
X X 0 X
By (TR3) and (TR4) it can be completed to a morphism of triangles, i.e. there

exists u : Y → X such that u ◦ u = id.
Similarly, one can show that v is a split epimorphism, i.e. there is a morphism
v : Z → Y such that v ◦ v = id.
5. Abelian categories vs. triangulated categories

As an application of the formal properties in the previous section we shall
compare the notions of abelian categories and triangulated categories.
Definition 5.1. An abelian category A is called semisimple if every short exact

sequence in A splits.
Example 5.2.
(i) Let R be a semisimple ring. Then the module categories R-Mod and R-mod
are semisimple. In particular, the category of vector spaces VecK over a
field K is semisimple.
(ii) The category Ab of abelian groups is not semisimple. For instance, the short
·2 ·1
exact sequence 0 → Z/2Z −→ Z/4Z −→ Z/2Z → 0 does not split.
The following result illustrates that the concepts of abelian and triangulated
categories overlap only slightly.
Theorem 5.3. Let T be a category which is triangulated and abelian. Then T

is semisimple.
f g
Proof. Let 0 → X −→ Y −→ Z → 0 be a short exact sequence in T . We
have to show that it splits; to this end it suffices to show that f is a section, i.e.
there exists a morphism f : Y → X such that f ◦ f = idX .
By (TR2) and (TR3), f can be embedded into a distinguished triangle
u f v
−1 V −→ X −→ Y −→ V .
The composition of consecutive morphisms in a distinguished triangle is always

zero by Proposition 4.1, in particular f ◦ u = 0. But f is a monomorphism in
T since it is the first map in a short exact sequence, hence u = 0. Thus we have
a distinguished triangle
f v u
X −→ Y −→ V −→ X
where u = 0. Now the triangle splits by Proposition 4.4.
We shall see in the next section that the homotopy category K(A) of com-
plexes over an additive category A is a triangulated category. This, together
with the preceding theorem, will then give a more structural explanation of the
earlier observation that K(Ab) is not abelian in Example 2.6, where we have
used an ad-hoc argument to show that morphisms do not necessarily have a
kernel.
6. The homotopy category of complexes is triangulated

Let A be an additive category, with corresponding category of complexes C(A)
and homotopy category K(A).
As discussed above, the homotopy category K(A) is in general not abelian,
even if A is abelian. We shall explain in this section how the homotopy category
K(A) becomes a triangulated category.
We first need an additive automorphism on K(A) which serves as transla-
tion functor. This functor can already be defined on the level of the category
C(A).
Definition 6.1. In C(A) we construct a translation functor = [1] by
shifting any complex one degree to the left. More precisely, for an object
X = (Xn , dnX )n∈Z in C(A) we set
X[1] := (X[1]n , dnX[1] )n∈Z with X[1]n = Xn−1 and dnX[1] = −dn−1
X
.
For a morphism of complexes f = (fn )n∈Z in C(A) we set
f [1] := (f [1]n )n∈Z where f [1]n = fn−1 .
Remark 6.2.
(i) The sign appearing in the differential of X[1] might look arbitrary; it
will become clear later when discussing the triangulated structure of the
homotopy category why this sign is needed.
(ii) The functor = [1] defined above is an additive functor and moreover
an automorphism of the category C(A).
(iii) Note that the above definitions are compatible with homotopies so we
have a well-defined induced functor = [1] on the homotopy category
K(A).
The next step for getting a triangulated structure on the homotopy category
is to find a suitable set of distinguished triangles. To this end, the following
construction of mapping cones is crucial.
Definition 6.3. Let f be a morphism between complexes X = (Xn , dnX ) and

Y = (Yn , dnY ). The mapping cone M(f ) is the complex in C(A) defined by

−dn−1
X
0
M(f )n = Xn−1 ⊕ Yn and dnM(f ) := .
fn−1 dnY
Remark 6.4.
(i) There are canonical morphisms in C(A) as follows
α(f ) : Y → M(f ), α(f )n := (0, idYn )
and
β(f ) : M(f ) → X[1], β(f )n := (idXn−1 , 0).
Note that β(f ) is a morphism of complexes because the differential in X[1]

carries a sign. From the above definitions we get a short exact sequence of
chain complexes
α(f ) β(f )
0 → Y −→ M(f ) −→ X[1] → 0.
(ii) Let f : X → Y be a morphism of complexes. The short exact sequence

α(f ) β(f )
0 → Y −→ M(f ) −→ X[1] → 0 splits (i.e. there is a morphism of com-
plexes σ : X[1] → M(f ) such that β(f ) ◦ σ = idX[1] ) if and only if f is
homotopic to zero. In fact, a splitting map is given by σ (x) := (x, −s(x))
where s is a homotopy map.
Example 6.5.
(i) For any complex X consider the zero map f : X → 0 to the zero complex.
Then the mapping cone is M(f ) = X[1]. On the other hand, the mapping
cone of g : 0 → Y is just M(g) = Y itself.
(ii) Let A and B be objects in A and view them as complexes XA and
XB concentrated in degree 0. Any morphism f : A → B in A induces
a morphism of complexes f : XA → XB . Its mapping cone is the
complex
f
. . . → 0 → A −→ B → 0 → . . .
where A is in degree 1 and B in degree 0.

(iii) Let X = (Xn , dnX ) be any complex in C(A). The mapping cone of the
identity morphism idX has degree n term equal to Xn−1 ⊕ Xn and differ-
ential

−dn−1X
0
: Xn−1 ⊕ Xn → Xn−2 ⊕ Xn−1 .
idXn−1 dnX
The identity morphism on the mapping cone M(id X ) is homotopic to zero,

0 idXn
via the map s = (sn )n∈Z where sn = . Thus, in the homotopy
0 0
category K(A) the identity idM(idX ) is equal to the zero map. As a conse-
quence, in the homotopy category, the mapping cone M(idX ) is isomorphic
to the zero complex.
It is easy to check that the morphisms α(f ) and β(f ) are also well-defined in
the homotopy category K(A) (i.e. independent on the choice of representatives
of the equivalence class of morphisms). This leads to the following definition.
Definition 6.6. A sequence of objects and morphisms in the homotopy category

K(A) of the form
f α(f ) β(f )
X −→ Y −→ M(f ) −→ X[1]
is called a standard triangle.

A distinguished triangle in K(A) is a triangle which is isomorphic (in K(A)!)
to a standard triangle.
With this class of distinguished triangles the homotopy category obtains a

triangulated structure as we shall show next. Due to the technical nature of the
axioms of a triangulated category, the proof that a certain additive category is
indeed triangulated is usually rather long, can be partly tedious and can still be
quite involved. In this introductory chapter we want to present such a proof at
least once in detail.
Theorem 6.7. Let A be an additive category. Then the homotopy category of

complexes K(A) is a triangulated category.
Proof. We have to show that with the above translation functor [1] and the set
of distinguished triangles just defined, the axioms (TR0)-(TR5) are satisfied.
The axioms (TR0) and (TR2) hold by Definition 6.6.
(TR1) From the mapping cone construction there is a standard triangle
idX
X −→ X −→ M(idX ) −→ X[1].
By Example 6.5 above, M(idX ) is isomorphic to the zero complex in the

homotopy category. Hence we indeed have a distinguished triangle
idX
X −→ X −→ 0 −→ X[1].
(TR3) Because the rotation property is compatible with isomorphisms of trian-

gles, it suffices to prove (TR3) for a standard triangle
f α(f ) β(f )
X −→ Y −→ M(f ) −→ X[1].
We shall show that the rotated triangle

α(f ) β(f ) −f [1]
Y −→ M(f ) −→ X[1] −→ Y [1]
is isomorphic in K(A) to the following standard triangle for α(f ),

α(f ) α(α(f )) β(α(f ))
Y −→ M(f ) −→ M(α(f )) −→ Y [1].
For constructing an isomorphism between the latter two triangles we take the
identity maps for the first, second and fourth entries. Moreover, we define
morphisms
φ = (φn ) : X[1] → M(α(f )) by setting φn = (−fn−1 , idXn−1 , 0)
and conversely
ψ = (ψn ) : M(α(f )) → X[1] by setting ψn = (0, idXn−1 , 0).
These yield morphisms of triangles since by definition β(α(f )) ◦ φ = −f [1],

and φ ◦ β(f ) ∼ α(α(f )) via the homotopy given by
⎛ ⎞
0 − id
⎝0 0 ⎠ : M(f )n = Xn−1 ⊕ Yn → M(α(f ))n+1 = Yn ⊕ Xn ⊕ Yn+1 .
0 0
Similarly, ψ is a morphism of triangles since β(f ) = ψ ◦ α(α(f )) by definition
and −f [1] ◦ ψ ∼ β(α(f )) via the homotopy (0, 0, − id) : M(α(f ))n → Y [1]n .
Finally, and most importantly for proving (TR3), the above morphisms are
isomorphisms in K(A) because we have ψ ◦ φ = idX[1] (by definition) and
φ ◦ ψ ∼ idM(α(f )) via the homotopy map
⎛ ⎞
0 0 − id
⎝0 0 0 ⎠ : M(α(f ))n → M(α(f ))n+1
0 0 0
(recall that M(α(f ))n = Yn−1 ⊕ Xn−1 ⊕ Yn ).
(TR4) Again it suffices to prove the axiom for standard triangles. By assump-
tion we have a diagram
u- α(u)
- M(u) β(u)
- X[1]
X Y
f g f [1]
? u ? α(u ) ?
X - Y - M(u ) β(u-) X [1]
where the left square commutes in K(A), i.e. there exist homotopy maps
Y
sn : Xn → Yn+1 such that gn un − un fn = dn+1 sn + sn−1 dnX for all n ∈ Z. For
completing the diagram to a morphism of triangles we define h = (hn )n∈Z :
M(u) → M(u ) by setting

fn−1 0
hn = : M(u)n = Xn−1 ⊕ Yn → M(u )n = Xn−1
⊕ Yn .
sn−1 gn
This is indeed a morphism of complexes because of the homotopy property of
s given above. Moreover, the completed diagram commutes since by definition
we have that h ◦ α(u) = α(u ) ◦ g and β(u ) ◦ h = f [1] ◦ β(u); note that these
are proper equalities, not only up to homotopy.
(TR5) Again it suffices to prove the octahedral axiom for standard trian-
gles. From the assumptions we already have the following part of the relevant
diagram
X
u -Y α(u)-
M(u)
β(u)-
X[1]
v
?
X
vu - Z α(vu)- M(vu) β(vu)
- X[1]
u u[1]
? ?
Y
v -Z α(v)-
M(v)
β(v)-
Y [1]
α(u) α(vu) α(u)[1]

? ? ?
M(u) M(vu) M(v) M(u)[1]
We now define the missing morphisms as follows. Let f = (fn ) : M(u) →
idXn−1 0
M(vu) be given in degree n by fn = and set g = (gn ) :
0 vn

un−1 0
M(vu) → M(v) to be given by gn = . Finally define h :
0 idZn
M(v) → M(u)[1] as the composition α(u)[1] ◦ β(v), i.e. it is given by the
0 0
matrix . Then it is easy to check from the definitions that all
idYn−1 0
squares in the completed diagram commute (not only up to homotopy).
For proving (TR5) it now remains to show that the bottom line
f g h
M(u) −→ M(vu) −→ M(v) −→ M(u)[1]
is a distinguished triangle in K(A). To this end we construct an isomorphism
to the standard triangle
f α(f ) β(f )
M(u) −→ M(vu) −→ M(f ) −→ M(u)[1].
Note that only the third entries in the triangles are different. So it suffices
to find morphisms σ = (σn ) : M(v) → M(f ) and τ = (τn ) : M(f ) → M(v)
leading to commutative diagrams (in K(A)!), i.e. we need that β(f ) ◦ σ = h,
h ◦ τ = β(f ), σ ◦ g = α(f ) and τ ◦ α(f ) = g, up to homotopy. Moreover, we
have to show that they are isomorphisms in the homotopy category. We set
⎛ ⎞
0 0

⎜ idYn−1 0 ⎟
σn := ⎜ ⎟ and τn := 0 idYn−1 un−1 0
.
⎝ 0 0 ⎠ 0 0 0 idZn
0 idZn
First, let us check that σ and τ give commutative diagrams. Directly from the
we get that τ ◦ α(f ) = g; in fact both are given in degree n by
definitions
un−1 0
the map : Xn−1 ⊕ Zn → Yn−1 ⊕ Zn . Also by definition we see
0 idZn

0 0
that β(f ) ◦ σ = h, both given by : Yn−1 ⊕ Zn → Xn−2 ⊕ Yn−1 .
idYn−1 0
The remaining commutativities will now only hold up to homotopy. Note that
α(f ) − σ ◦ g : M(vu) → M(f ) is given in degree n by
⎛ ⎞
0 0
⎜ −un−1 0 ⎟
⎜ ⎟ : Xn−1 ⊕ Zn → Xn−2 ⊕ Yn−1 ⊕ Xn−1 ⊕ Zn .
⎝ idX 0⎠
n−1
0 0
We claim that α(f ) − σ ◦ g is homotopic to zero, i.e. α(f ) = σ ◦ g in K(A).
In fact, a homotopy map s = (sn ) where sn : M(vu)n → M(f )n+1 is given by
⎛ ⎞
idXn−1 0
⎜ 0 0⎟
⎜ ⎟ : Xn−1 ⊕ Zn → Xn−1 ⊕ Yn ⊕ Xn ⊕ Zn+1 .
⎝ 0 0⎠
0 0
For verifying the details recall that the differential of the mapping cone M(f )
is given by
⎛ X ⎞
dn−2 0 0 0
⎜ −un−2 −dn−1 Y
0 0 ⎟
dnM(f ) = ⎜
⎝ idX
⎟.
n−2
0 −d X
0 ⎠
n−1
0 vn−1 (vu)n−1 dnZ
Finally, consider β(f ) − h ◦ τ : M(f ) → M(u)[1] which in degree n is given
by

idXn−2 0 0 0
: Xn−2 ⊕ Yn−1 ⊕ Xn−1 ⊕ Zn → Xn−2 ⊕ Yn−1 .
0 0 −un−1 0
This can be seen to be homotopic to zero by using the homotopy map s = (sn )
where

0 0 idXn−1 0
sn = : Xn−2 ⊕ Yn−1 ⊕ Xn−1 ⊕ Zn → Xn−1 ⊕ Yn .
0 0 0 0
For the straightforward verification again use the differential of M(f ) as given
above.
For completing the proof it now remains to show that σ and τ are iso-
morphisms in the homotopy category. We have τ ◦ σ = idM(v) by definition.
Conversely, the composition σ ◦ τ is in degree n given by
⎛ ⎞
0 0 0 0
⎜ 0 idYn−1 un−1 0 ⎟
⎜ ⎟
⎝0 0 0 0 ⎠
0 0 0 idZn
If we then define homotopy maps sn : M(f )n → M(f )n+1 by setting
⎛ ⎞
0 0 − idXn−1 0
⎜0 0 0 0⎟
sn := ⎜
⎝0 0
⎟
0 0⎠
0 0 0 0
M(f ) M(f )
then we have σ ◦ τ − idM(f ) = dn+1 ◦ sn + sn−1 ◦ dn which is easily
checked using the differential of M(f ) as given above.
Thus σ ◦ τ = idM(f ) in the homotopy category K(A) and we have proved
the octahedral axiom for K(A).
Remark 6.8. We have seen that for every standard triangle

f α(f ) β(f )
X −→ Y −→ M(f ) −→ X[1]
in K(A) there is a corresponding short exact sequence

α(f ) β(f )
0 → Y −→ M(f ) −→ X[1] → 0
in C(A). On the other hand, it is not true that any short exact sequence in C(A)
would lead to a distinguished triangle in the homotopy category K(A).
As an example, consider the short exact sequence of abelian groups
·2 ·1
0 → Z/2Z −→ Z/4Z −→ Z/2Z → 0,
and consider the abelian groups as complexes concentrated in degree 0. There
is no corresponding distinguished triangle
·2 ·1 w
Z/2Z −→ Z/4Z −→ Z/2Z −→ Z/2Z[1].
in K(Ab). In fact, suppose for a contradiction that such a distinguished triangle
existed. The morphisms in K(Ab) are just equivalence classes of morphisms of
complexes modulo homotopy. But since Z/2Z is a complex concentrated in a
single degree, there are no nonzero morphisms Z/2Z → Z/2Z[1] to its shifted
version. Thus we must have w = 0.
By Proposition 4.4 a triangle with a zero map is a split triangle. Hence
Z/4Z ∼ = Z/2Z ⊕ Z/2Z in K(Ab), i.e. there must exist a homotopy equiva-
lence between these complexes. However, all these complexes are complexes
concentrated in a single degree, hence there are no nonzero homotopy maps.
So the above isomorphism would have to be an isomorphism already in Ab
which is impossible, a contradiction.
We shall later see that this phenomenon disappears when passing from
K(A) to the derived category. There every short exact sequence does lead to a
distinguished triangle; see Section 7.6 below for details.
7. Derived categories
A very important class of triangulated categories is formed by derived cat-
egories. They occur frequently in many different areas of mathematics and
have found numerous applications. In this section we shall provide the relevant
constructions leading from the homotopy category to the derived category.
7.1. Homology and quasi-isomorphisms

In this short section we shall introduce the notion of quasi-isomorphism which
is fundamental for derived categories.
Although one could set up a homology theory in a categorical manner in
every abelian category we shall restrict from now on to categories of modules
and to complexes over them. This considerably simplifies the presentation in

certain parts of this section since we can then use element-wise arguments and
hence avoid technical overload which might obscure the fundamental ideas
underlying the definition of a derived category.
For the remainder of this section we let A be a category of modules over a
ring.
Definition 7.1.
(i) Let X = (Xn , dnX ) be a complex in C(A). The n-th homology of the com-
plex X is defined as the following object from A,
Hn (X) := ker dnX / im dn+1

X
(where the kernel and the image are the usual set-theoretic kernel and
image, respectively).
(ii) The complex X is called exact if Hn (X) = 0 for all n ∈ Z.
(iii) Let f : X → Y be a morphism of complexes. We define an induced map
on the level of homology by setting
Hn (f ) : Hn (X) → Hn (Y ), x + im dn+1
X

→ fn (x) + im dn+1
Y
.
In this way we get homology functors Hn : C(A) → A where n ∈ Z.
Remark 7.2. Note that the induced map on homology is well-defined; in fact
let x = dn+1
X
(x ) ∈ im dn+1
X
; then
fn (x ) = fn (dn+1
X
(x )) = dn+1
Y
(fn+1 (x )) ∈ im dn+1
Y
.
Proposition 7.3. Let f, g : X → Y be morphisms in C(A) which are homo-

topic. Then they induce the same map in homology, i.e. Hn (f ) = Hn (g) for all
n ∈ Z.
As a consequence, the homology functors on C(A) induce well-defined
homology functors on the homotopy category K(A).
Proof. By assumption there is a homotopy map s = (sn )n∈Z such that fn −

gn = dn+1
Y
sn + sn−1 dnX for all n ∈ Z. Let x + im dn+1
X
∈ Hn (X), in particular
x ∈ ker dn . Then it follows that
X
Hn (f )(x + im dn+1
X
) = fn (x) + im dn+1
Y
= (dn+1
Y
sn + sn−1 dnX + gn )(x) + im dn+1
Y
= gn (x) + im dn+1
Y
= Hn (g)(x + im dn+1
X
).
Hence H (f ) = H (g).
Definition 7.4. A morphism f : X → Y of complexes in C(A) is called a quasi-

isomorphism if it induces isomorphisms in homology, i.e. Hn (f ) : Hn (X) →
Hn (Y ) are isomorphisms for all n ∈ Z.
Example 7.5.
(i) (Projective resolutions) As we are restricting in this section to categories R-
Mod of modules over a ring any object X in A has a projective resolution,
i.e. a sequence
. . . → P2 → P1 → P0 → 0
where all Pi are projective objects, together with a morphism

: P0 → X
such that the following augmented sequence is exact

. . . → P2 → P1 → P0 → X → 0.
This gives rise to a morphism of complexes, also denoted

: P → X,
... → P2 → P1 → P0 → 0
↓ ↓ ↓
... → 0 → 0 → X → 0
where X is supposed to be in degree 0. Then
is a quasi-isomorphism.
In fact, in non-zero degrees both complexes have zero homology, and in
degree 0 we have isomorphisms H0 (P ) ∼ =X∼ = H0 (X) induced by
.
(ii) (Injective resolutions) Dually, every object has an injective resolution, i.e.
there is a sequence
0 → I0 → I−1 → I−2 → . . .
where all Ij are injective objects, and a morphism ι : X → I0 such that the
following augmented sequence is exact
ι
0 → X −→ I0 → I−1 → I−2 → . . .
This also induces a morphism of complexes ι which is a quasi-isomorphism.
Our next goal is to characterize quasi-isomorphisms in terms of mapping

cones. To this end we shall use the following standard result on long exact
sequences; for a proof we refer for instance to Weibel’s book [17, section 1.3].
f g
Proposition 7.6. Let 0 → X −→ Y −→ Z → 0 be a short exact sequence
of complexes in C(A). Then there are connecting morphisms δn : Hn (Z) →
Hn−1 (X) giving rise to the following long exact homology sequence
Hn (g) δn+1 Hn (f ) Hn (g) δn Hn (g)
. . . −→ Hn+1 (Z) −→ Hn (X) −→ Hn (Y ) −→ Hn (Z) −→ Hn−1 (X) −→ . . .
For the definition of mapping cones recall Definition 6.3.

Proposition 7.7. In C(A) a morphism f : X → Y is a quasi-isomorphism if
and only if the mapping cone complex M(f ) is exact.
Proof. By Remark 6.4 we have an exact sequence of complexes
α(f ) β(f )
0 → Y −→ M(f ) −→ X[1] → 0.
The corresponding long exact homology sequence has the form
δn+1 Hn (α(f )) Hn (β(f )) δn
. . . Hn+1 (X[1]) −→ Hn (Y ) −→ Hn (M(f )) −→ Hn (X[1]) −→ Hn−1 (Y ) . . .
But Hn (X[1]) can be identified with Hn−1 (X) for all n ∈ Z and it can be checked
that then in our situation δn = Hn−1 (f ), so the above long exact sequence takes
the form
Hn (f ) Hn (α(f )) Hn (β(f )) Hn−1 (f )
. . . Hn (X) −→ Hn (Y ) −→ Hn (M(f )) −→ Hn−1 (X) −→ Hn−1 (Y ) . . .
For necessity, suppose that f is a quasi-isomorphism. Then Hn (f ) are iso-
morphisms by assumption, hence by exactness we have Hn (α(f )) = 0 and
Hn (β(f )) = 0 for all n ∈ Z. But then again by exactness we deduce that
Hn (M(f )) = 0 for all n ∈ Z, i.e. the mapping cone M(f ) is exact.
Conversely, suppose that M(f ) is exact. Then the long exact sequence takes
the form
Hn (f ) Hn−1 (f )
. . . 0 → Hn (X) −→ Hn (Y ) −→ 0 −→ Hn−1 (X) −→ Hn−1 (Y ) → 0 . . .
from which it immediately follows by exactness that Hn (f ) is an isomorphism
for all n ∈ Z, i.e. f is a quasi-isomorphism.
f g h
Remark 7.8. Let X −→ Y −→ Z −→ X[1] be a distinguished triangle in
K(A). Then it follows from the previous proposition that f is a quasi-
isomorphism if and only if Z is exact (i.e. Hn (Z) = 0 for all n ∈ Z).
7.2. Localisation of categories

Derived categories of abelian categories are obtained from the homotopy cate-
gories of complexes by localising with respect to quasi-isomorphisms, i.e. by
a formal process inverting all quasi-isomorphisms.
We shall not aim in this introductory chapter to provide a general account of
localisation of categories. For a thorough treatment of this topic see H. Krause’s
article in this volume [10].
Instead we shall concentrate here on the special case leading from the
homotopy category K(A) of complexes to the derived category D(A). We first
want to give an elementary construction of a derived category, following the

approach in the book by Gelfand and Manin [4, III.2]. In the following sections
we shall also give alternative equivalent descriptions of the derived category
which are perhaps more common and more suitable for explicit computations.
Remark 7.9. From now on we are following a time-honoured tradition by
ignoring some set-theoretical issues.
Theorem 7.10. Let A be an abelian category. Then there exists a category
D(A), called the derived category of A, and a functor L : K(A) → D(A)
satisfying the following properties:
(L1) For every quasi-isomorphism q in K(A), L(q) is an isomorphism in
D(A).
(L2) Every functor F : K(A) → D (where D is any category), having the
property that quasi-isomorphisms are mapped to isomorphisms, factors
uniquely through L.
Property (L2) implies in particular that the category D(A), if it exists, is
unique up to equivalence of categories.
Proof. The objects in D(A) are defined to be the same as in K(A), i.e. complexes
over A. But the morphisms have to be changed in order for quasi-isomorphisms
to become isomorphisms. For each quasi-isomorphism q in K(A) we introduce
a formal variable q −1 . We then consider ‘words’ in f ’s and q −1 ’s, i.e. formal
compositions of the form
(∗) = f1 ◦ q1−1 ◦ f2 ◦ q2−1 ◦ . . . ◦ fr ◦ qr−1
where r ∈ N0 , the fi are morphisms and the qj are quasi-isomorphisms in
K(A). This has to be read so that some fi or some qj−1 can be the identity and
then can be deleted, i.e. consecutive subexpressions fi ◦ fi+1 or qj−1 ◦ qj−1+1 are
also allowed in (*).
As usual we read compositions from right to left; so if f1 : X1 → Y1 then Y1
is called the end point of (∗) and if qr : Xr → Yr then Yr is called the starting
point of (∗). The length of (∗) is the total number of fi ’s and qj−1 ’s occurring.
For each object X there is an empty expression of length 0 representing the
identity on X.
We call two such expressions equivalent if they have the same starting and
end point and if one can be obtained from the other by a sequence of the
following operations
(i) for any composable morphisms f, g in K(A) replace f ◦ g by their com-
position (f ◦ g);
(i’) for any composable quasi-isomorphisms q, r in K(A) replace q −1 ◦ r −1

by (r ◦ q)−1 ;
(ii) for any quasi-isomorphism q in K(A) replace q ◦ q −1 or q −1 ◦ q by id.
The morphisms in D(A) are defined as equivalence classes of expressions

of the form (∗). The composition of morphisms is induced by concatenating
expressions of the form (∗) if the starting point of the second matches the end
point of the first, and zero otherwise.
The crucial localisation functor L is now defined as follows: on objects, L
is just the identity; a morphism f in K(A) is sent by L to its equivalence class
in D(A).
In particular, if q is a quasi-isomorphism in K(A) then L(q) becomes invert-
ible in D(A) with inverse q −1 (because q ◦ q −1 is equivalent to the empty
expression of length 0, representing the identity). Thus, axiom (L1) is satisfied.
For proving axiom (L2) let a functor F : K(A) → D be given (where D is
any category) which sends quasi-isomorphisms to isomorphisms. We need to
define a functor G : D(A) → D such that G ◦ L = F . First we note that there
is at most one possibility to define such a functor, namely setting G(X) = F (X)
on objects, defining G(f ) = F (f ) for morphisms f in K(A) and G(q −1 ) =
F (q)−1 for quasi-isomorphisms q in K(A) (and then extending G to arbitrary
compositions, in particular G(id) = id for the empty composition). Note that
the latter makes sense since F (q) is an isomorphism by assumption.
It only remains to check that this functor is well-defined, i.e. compatible
with the equivalence relation defining morphisms in D(A). For instance, for
part (ii) of the above equivalence relation we have in D(A) that
G(q ◦ q −1 ) = G(q) ◦ G(q −1 ) = F (q) ◦ F (q)−1 = id = G(id)
showing well-definedness. The other parts also follow easily from the
definition.
7.3. Morphisms in the derived category

The above description of morphisms in the derived category as equivalence
classes of expressions of the form
(∗) = f1 ◦ q1−1 ◦ f2 ◦ q2−1 ◦ . . . ◦ ◦fr ◦ qr−1
is pretty inconvenient. We shall describe in this section a ‘calculus of fractions’

which will lead to a simpler description of the morphisms in the derived cate-
gory. To this end we shall make use of certain useful properties of the class of
quasi-isomorphisms in the homotopy category.
Lemma 7.11. Let A be an abelian category. The class Q of quasi-isomorphisms

in the homotopy category K(A) satisfies the following properties:
(Q1) For every object X in K(A) the identity idX is in Q.

(Q2) Q is closed under composition.
(Q3) (Ore condition) Given a quasi-isomorphism q ∈ Q and a morphism f in
K(A) (with the same target) then there exist an object W , a morphism
g and a quasi-isomorphism t ∈ Q such that the following diagram is
commutative.
g
W⏐ −→ ⏐Z
t q
f
X −→ Y
Similarly, given a quasi-isomorphism q ∈ Q and a morphism f in K(A)

(with the same range) then there exist an object V , a morphism h
and a quasi-isomorphism r ∈ Q such that the following diagram is
commutative.
f
Y⏐ −→ ⏐
X
q r
h
X −→ V
(Q4) For any morphisms f, g : X → Y in K(A) the following are equivalent:

(i) There exists a quasi-isomorphism q : Y → Y in Q (for some Y )
such that q ◦ f = q ◦ g.
(ii) There exists a quasi-isomorphism t : X → X in Q (for some X )
such that f ◦ t = g ◦ t.
Remark 7.12. The class of quasi-isomorphisms does not in general satisfy the
conditions of the preceding lemma already in C(A); it is crucial first to pass to
the homotopy category. In fact, in the proof below we shall make heavy use of
the triangulated structure of the homotopy category (which has been proven in
Theorem 6.7).
Proof. (Q1) and (Q2) are clear.

For (Q3) we have given a morphism f and a quasi-isomorphism q. By axiom
(TR2) for the triangulated category K(A) there exists a distinguished triangle
q u v
Z −→ Y −→ U −→ Z[1]. Similarly, considering uf : X → U there is a dis-
t uf w
tinguished triangle W −→ X −→ U −→ W [1]. Applying axioms (TR4) and
(TR3) we can deduce the existence of the morphism g (and g[1]) in the follow-
ing commutative diagram
t- uf- w-
W X U W [1]
g f id g[1]
? ? ? ?
q- u- v-
Z Y U Z[1]
Since q is a quasi-isomorphism by assumption, the long exact homology
sequence applied to the bottom row yields that Hn (U ) = 0 for all n ∈ Z. And
then the long exact homology sequence for the top row implies that t must be
a quasi-isomorphism, as desired (cf. Remark 7.8).
The symmetrical second claim in (Q3) is shown similarly.
Finally, let us prove (Q4). We will prove the direction (i)⇒(ii), the con-
verse is proved similarly. For simplicity, set h := f − g, thus q ◦ h = 0
by assumption. By (TR2) and (TR3) there exists a distinguished triangle
u q w
Z −→ Y −→ Y −→ Z[1]. Since q is a quasi-isomorphism, Z[1] and hence Z
is exact (cf. Remark 7.8). By (TR1), (TR3) and (TR4) there exists a morphism
v making the following diagram commutative
id- - 0 - X[1]
X X
v h v[1]
? ? ? ?
u- q - w-
Z Y Y Z[1]
Now again by (TR2) and (TR3) v can be embedded in a distinguished triangle
t v
X −→ X −→ Z −→ X [1]. Here t is a quasi-isomorphism since Z is exact
(cf. Remark 7.8). Moreover, v ◦ t = 0 since the composition of any consecutive
maps in a distinguished triangle vanishes. It follows that h ◦ t = u ◦ v ◦ t = 0,
i.e. f ◦ t = g ◦ t, as desired.
The properties satisfied by the family Q of quasi-isomorphisms in K(A) has

useful consequences for the description of morphisms in the derived category
D(A). As described above, a morphism in D(A) is an equivalence class of an
expression of the form
(∗) = f1 ◦ q1−1 ◦ f2 ◦ q2−1 ◦ . . . ◦ ◦fr ◦ qr−1
where fi are morphisms in K(A) and qi are quasi-isomorphisms in Q.
Property (Q3) above states that q −1 ◦ f = g ◦ t −1 for some morphism g and
quasi-isomorphism t ∈ Q, and f ◦ q −1 = r −1 ◦ h with r ∈ Q, respectively.
This means that in the above expression (*) we can move all ‘denominators’
qi to the right (or to the left). This means that any morphism in the derived
category can be represented by an expression of the form f ◦ q −1 with a quasi-
isomorphism q and a morphism f . This can be conveniently visualised as a
‘roof’.
X
q @f
R
@
X Y
We shall use this description frequently in the sequel and hence want to make
this more precise. Again, we follow the approach in the book by Gelfand and
Manin [4, section III.2]. We shall define a category
D(A) where morphisms are
represented by such roofs, and then show that this category is indeed equivalent
to the derived category D(A) introduced in Theorem 7.10.
For computing with these roofs we need to introduce a suitable notion of
equivalence for roofs. Two roofs (q, f ) and (t, g) are called equivalent if there
exists another roof (r, h) making the following diagram commutative.
X
r @h
R
@
X X
q P P g
PPP @

)t f Pq
P R
@
X Y
We leave it to the reader to verify that this indeed defines an equivalence

relation. The non-obvious property is transitivity, see [4, Lemma III.2.8] for
a detailed proof; actually, in the proof of transitivity the property (Q4) of
Lemma 7.11 is used.
For the composition of roofs one makes use of the Ore condition (Q3) above.
Namely, given roofs (q, f ) : X → Y and (t, g) : Y → Z we can find by (Q3)
an object W and a roof (t , g ) making the following diagram commutative
W
t
@g

R
@
X Y
q @f t @g
R
@ R
@
X Y Z
The composition of roofs (t, g) ◦ (q, f ) is then defined to be the equivalence

class represented by the roof (q ◦ t , g ◦ g ) : X → Z. It is not difficult to
verify that this is well-defined, i.e. independent of the representatives of the
roofs involved.
Note that the identity morphism for an object X is represented by the roof
(idX , idX ).
The category
D(A) is defined as having the same objects as D(A) (and hence
as the homotopy category K(A)), namely complexes over A.
The morphisms in D(A) are defined to be the equivalence classes of roofs,
with the above composition.
Proposition 7.13. Let A be an abelian category. Then the category D(A)
satisfies the universal property of the derived category D(A) given in Theorem
7.10. In particular, the categories D(A) and D(A) are equivalent.
Proof. We first define a functor L : K(A) → D(A) as the identity on objects
and on morphisms by sending f to the roof (id, f ). Clearly, L maps the identity
to the identity. As for composition of morphisms, a composition g ◦ f is on the
one hand sent by L to the roof (id, g ◦ f ); on the other hand the composition
◦ L(f
L(g) ) is given by the roof obtained from the commutative diagram
X
id @f
R
@
X Y
id @f id @g
R
@ R
@
X Y Z
◦ f ) = L(g)
Thus, L(g ◦ L(f ) and L is indeed a functor.
Now it remains to prove that the category D(A), together with the functor

L, satisfies the properties (L1) and (L2) from Theorem 7.10.
For (L1), any quasi-isomorphism q in K(A) is mapped to the roof (id, q). It
is immediate from the above composition of morphisms that (q, id) ◦ (id, q) =
(id, id), and that (id, q) ◦ (q, id) = (q, q); but the latter roof is equivalent
to (id, id). Thus, L maps quasi-isomorphisms to isomorphisms and (L1) is
satisfied.
For proving (L2), let F : K(A) → D (D any category) be a functor which
maps quasi-isomorphisms to isomorphisms. We have to show that there is a
unique functor F : D(A) → D such that F ◦ L = F.
We first deal with uniqueness. On objects X, the only choice is F (X) =
F (X) since L is the identity on objects. Now consider a morphism in D(A),

represented by a roof (q, f ). In D(A) we have that
= (q, f ) ◦ (id, q) = (id, f ) = L(f
(q, f ) ◦ L(q) ).
has to be a functor with F
Since F ◦ L
= F we can deduce that
(L(f
F (f ) = F )) = F ((q, f ) ◦ L(q))
((q, f )) ◦ F
=F (L(q))

=F((q, f )) ◦ F (q).
on
By assumption, F (q) is an isomorphism, so the only possibility to define F
morphisms is to set
((q, f )) = F (f ) ◦ F (q)−1 .
F
Hence, the functor F, if it exists, is unique.

For existence, we actually define F by the properties just exhibited, i.e.
(X) = F (X) on objects and F
F ((q, f )) = F (f ) ◦ F (q)−1 on morphisms. Of
course, we now have to prove that this indeed defines a functor.
We claim that the definition is well-defined, i.e. independent of the choice
of the representative. In fact, let (q, f ) and (t, g) be equivalent roofs, i.e. we
have a commutative diagram of the form
X
r @h
R
@
X X
q P P g
PPP @

)t f Pq
P R
@
X Y
where r is a quasi-isomorphism. Note that since r, q, t are quasi-isomorphisms

and q ◦ r = t ◦ h, also h must be a quasi-isomorphism. Then we get from
the functoriality of F and the assumption that F sends quasi-isomorphisms to
isomorphisms that
((q, f )) = F (f ) ◦ F (q)−1 = F (f ) ◦ F (r) ◦ F (r)−1 ◦ F (q)−1
F
= F (f ◦ r) ◦ (F (q ◦ r))−1 = F (g ◦ h) ◦ (F (t ◦ h))−1
((t, g)).
= F (g) ◦ F (h) ◦ F (h)−1 ◦ F (t)−1 = F (g) ◦ F (t)−1 = F
By definition, F maps identity morphisms (id, id) in D(A) to identity

morphisms in D. Finally, consider a composition (t, g) ◦ (q, f ) in
D(A);

this is represented by a roof (q ◦ t , g ◦ g ) coming from a commutative
diagram
W
t
@g

R
@
X Y
q @f t @g
R
@ R
@
X Y Z
Since f ◦ t = t ◦ g we get F (f ) ◦ F (t ) = F (t) ◦ F (g ) and since F sends

quasi-isomorphisms to isomorphisms F (t)−1 ◦ F (f ) = F (g ) ◦ F (t )−1 . This
implies that
((t, g) ◦ (q, f )) = F
F ((q ◦ t , g ◦ g )) = F (g ◦ g ) ◦ F (q ◦ t )−1
= F (g) ◦ F (g ) ◦ F (t )−1 ◦ F (q)−1
= F (g) ◦ F (t)−1 ◦ F (f ) ◦ F (q)−1
=F((t, g)) ◦ F
((q, f )).
Thus F is indeed a functor and this completes the proof of the universal property
(L2).
Remark 7.14. In the sequel we shall denote the derived category exclusively
by D(A) even if we usually use the more convenient equivalent version
D(A)
just described.
Proposition 7.15. Let A be an abelian category. Then the derived category

D(A) is an additive category.
Proof. Following Definition 1.1 we have to show the properties (A1), (A2) and
(A3).
(A1) We first describe addition of morphisms. Let two morphisms F, G
from X to Y be represented by roofs (q, f ) and (q , f ). By the Ore condition
(Q3) there exists an object W , a morphism g and a quasi-isomorphism t making
the following diagram commutative
g
W⏐ −→ X
⏐
t q
q
X −→ X
Since q, q and t are quasi-isomorphisms, also g must be a quasi-isomorphism.

From the definition of equivalence it is easy to check that the roof (q, f ) is
equivalent to the roof (q ◦ g, f ◦ g), and that (q , f ) is equivalent to the roof
(q ◦ t, f ◦ t) = (q ◦ g, f ◦ t). Thus we have found a ‘common denominator’
and can set F + G to be the roof represented by (q ◦ g, f ◦ g + f ◦ t).
We leave it to the reader to verify that this addition is well-defined (i.e.
independent of the representatives) and that the addition of morphisms is
bilinear.
Note that in the derived category there is for any objects X, Y a zero mor-
phism in D(A) which is represented by the roof (idX , 0X,Y ) where 0X,Y is the
zero morphism of complexes from X to Y .
(A2) The zero object in D(A) is the zero complex (i.e. it is the same
zero object as in the homotopy category). We have to show that for every
object X the morphism sets HomD(A) (X, 0) and HomD(A) (0, X) contain only
the morphism represented by the roof (idX , 0) and (0, idX ), respectively. In
fact, any morphism from X to the zero complex is represented by a roof
(q, 0) where q : Z → X is a quasi-isomorphism. But it easily follows from
the definition that the roof (q, 0) is equivalent to (idX , 0), thus HomD(A) (X, 0)
contains precisely one element. The assertion for HomD(A) (0, X) is shown
similarly.
(A3) For the coproduct of two objects X and Y in D(A) one uses the
image of the coproduct X ⊕ Y in K(A) under the localisation functor L
(which is the identity on objects and maps a morphism f in K(A) to the roof
(id, f ) in D(A)). The corresponding maps L(ιX ) : X → X ⊕ Y and L(ιY ) :
X → X ⊕ Y are given by the roofs (id, ιX ) and (id, ιY ), respectively, where
ιX and ιY are the embeddings (or more precisely, their equivalence classes in
K(A)).
We have to show that the universal property (A3) is satisfied. So let f˜X and
˜
fY be arbitrary morphisms in D(A) from X and Y to some object Z. They are
represented by roofs of the form
X̃ Ỹ
q @fX and t @fY
R
@ R
@
X Z Y Z
The required morphism f˜ from X ⊕ Y to Z can then be defined as being rep-

resented by the roof (q ⊕ t, (fX , fY )) (where the notation ⊕ on maps between
direct sums of complexes means componentwise application). Then indeed we
have that the composition f˜ ◦ L(ιX ) in D(A) is the roof (q, fX ), as can be seen
from the diagram
X̃
q @ ιX̃
@
R
@
X X̃ ⊕ Ỹ
id @
ιX q ⊕ t @ (fX , fY )
@
R
@ @ R
@
X X⊕Y Z
A similar diagram shows that f˜ ◦ L(ιY ) = (t, fY ) in D(A).

We leave it as an exercise to show that the morphism f˜ with these properties
is actually unique, as required in (A3).
7.4. Derived categories are triangulated

Recall that the derived category has been obtained by localising the homotopy
category with respect to the class of quasi-isomorphisms. In particular, there is
a functor L : K(A) → D(A) sending quasi-isomorphisms in K(A) to isomor-
phisms in D(A) (and satisfying a universal property). We have seen earlier that
the homotopy category is triangulated, with distinguished triangles being the
triangles isomorphic in K(A) to the standard triangles coming from mapping
f α(f ) β(f )
cones X −→ Y −→ M(f ) −→ X[1].
For obtaining a triangulated structure on the derived category the idea is to
transport the triangulated structure on the homotopy category via the localisa-
tion functor L.
Definition 7.16. The translation functor on D(A) is defined as the shift [1] on
objects, and for a morphism F in D(A) represented by a roof (q, f ) we set
F [1] to be the equivalence class of the roof (q[1], f [1]).
A triangle in D(A) is a distinguished triangle if it is isomorphic (in D(A)!)
to the image of a distinguished triangle from K(A) under the localisation
functor L.
Remark 7.17. When passing from K(A) to D(A) all quasi-isomorphisms

become isomorphisms, i.e. there are ‘more’ isomorphisms in D(A) than in
K(A). This in turn means that in D(A) it is easier for a triangle to become
isomorphic to a standard triangle than in K(A), i.e. the derived category contains
‘more’ distinguished triangles than the homotopy category.
As a crucial observation we shall see in the next section that the derived
category has the property that every short exact sequence of complexes in C(A)
gives rise to a corresponding distinguished triangle in the derived category.
This is not yet the case in the homotopy category, see Remark 6.8 above for an
example.
With the above definitions one can then show the main structural property
of derived categories. Unfortunately, the proof of the axioms for a triangulated
category will become very technical so that we shall refrain from providing a
proof in this introductory chapter.
Theorem 7.18. Let A be an abelian category. Then the derived category D(A)
is triangulated.
7.5. Comparing morphisms

Let A be an abelian category. We have now constructed various categories from
A:
A
→ C(A)
→ K(A)
→ D(A)
abelian abelian triangulated triangulated
Proposition 7.19. We have the following implications for a morphism f in

C(A).
f = 0 in C(A) ⇒ f = 0 in K(A) ⇒ f = 0 in D(A) ⇒ Hn (f )

= 0 for all n ∈ Z
Proof. The first two implications are obvious. For the third, f : X → Y as
a morphism in D(A) is represented by the roof (idX , f ). For f being 0 in
D(A) means being equivalent to the roof (idX , 0X,Y ), i.e. in K(A) there is a
commutative diagram
Z
r @h
R
@
X X
id PP 0
PPP @
id
) f Pq@
PR
X Y
with a quasi-isomorphism r. By commutativity, f ◦ r = 0 (in K(A)) and

passing to homology we get Hn (f ) ◦ Hn (r) = Hn (f ◦ r) = Hn (0) = 0 for all
n ∈ Z. But r is a quasi-isomorphism, i.e. all Hn (r) are isomorphisms, thus we
conclude that Hn (f ) = 0 for all n ∈ Z.
Remark 7.20.
(1) Note that a morphism f in K(A) becomes zero in D(A) if and only if there
exists a quasi-isomorphism r such that f ◦ r is homotopic to zero.
(2) All implications given in Proposition 7.19 are strict. Let us give examples
for each case. We consider the category A = Ab.
For the first implication, consider the following morphism of complexes
... → 0 → 0 −→ Z⏐ → 0 → ...
id
id
... → 0 → Z −→ Z → 0 → ...
Clearly, this morphism is zero in K(A), but nonzero in C(A).

For the second implication, consider the identity map on the (exact)
complex
·2 π
0 → Z −→ Z −→ Z/2Z → 0.
This morphism is not homotopic to zero (i.e. nonzero in K(A)) because
HomZ (Z/2Z, Z) = 0 and hence the identity on Z/2Z can not factor through
a homotopy. However, it is zero in D(A) because we can find a quasi-
isomorphism r such that f ◦ r is homotopic to zero. In fact, let r be the
morphism of complexes
id
0 → 0 → Z⏐ −→ ⏐Z → 0
id π
·2 π
0 → Z −→ Z −→ Z/2Z → 0
Since both complexes are exact, r is a quasi-isomorphism. Moreover,
f ◦ r is homotopic to zero (a homotopy map is given by 0 and id in
the two relevant degrees). This implies that f is zero when considered as
a morphism in D(A).
For the third implication we consider the morphism f : X → Y of
complexes given as follows
·2
0 → Z⏐ −→ ⏐Z → 0
id π
·2
0 → Z −→ Z/3Z → 0
The homology of X is given by H0 (X) = Z/2Z and H1 (X) = 0, whereas
H0 (Y ) = 0 and H1 (Y ) = 3Z (and all other being zero). In particular,
Hn (f ) = 0 for all n ∈ Z.
However, we claim that f is nonzero in the derived category. Suppose
for a contradiction that f = 0 in D(A), i.e. there exist a complex R =
(Rn , dnR ) and a quasi-isomorphism r : R → X such that f ◦ r : R → Y is
homotopic to zero. Since r is a quasi-isomorphism, we have that Hn (R) ∼ =
Hn (X) = 0 for n = 0 and H0 (R) = H0 (X) ∼ = Z/2Z. Choose a generator
of H0 (R), i.e. z0 ∈ ker d0R \ im d1R . Since r is a quasi-isomorphism, r0 (z0 )
must not be in the image of d1X , i.e. r0 (z0 ) ∈ 2Z. On the other hand, f ◦ r
is homotopic to zero, thus there exist homotopy maps s0 : R0 → Z and
s−1 : R−1 → Z/3Z such that (f ◦ r)0 = π ◦ r0 = 2s0 + s−1 ◦ d0R . Applied
to the generator z0 (which is in the kernel of d0R ) this yields
(π ◦ r0 )(z0 ) = (2s0 + s−1 ◦ d0R )(z0 ) = 2s0 (z0 ) ∈ 2(Z/3Z).
But then also r0 (z0 ) ∈ 2Z, a contradiction to the earlier conclusion.
Hence there is no such quasi-isomorphism r, i.e. f = 0 in D(A).
7.6. Short exact sequences vs. triangles

In this section we shall explain the crucial observation that a short exact
sequence in C(A) induces a distinguished triangle in D(A). Recall that we
have seen earlier that this does not yet happen in the homotopy category K(A)
(cf. Remark 6.8).
In this subsection we will again use our assumption that the abelian category
A is a category of modules over a ring, which allows us to define maps on
elements.
7.6.1. Mapping cylinders

Let f : X → Y be a morphism of complexes in C(A). The mapping cylinder
of f is the complex Cyl(f ) having degree n part equal to Xn ⊕ Xn−1 ⊕ Yn and
the differential is given by
dnCyl(f ) (x, x , y) := (dnX x − x , −dn−1

X
x , fn−1 (x ) + dnY y).
In perhaps more convenient matrix notation,

⎛ X ⎞
dn − idXn−1 0
dnCyl(f ) = ⎝ 0 −dn−1
X
0 ⎠.
0 fn−1 dnY
Cyl(f ) Cyl(f )
It is now easily checked that Cyl(f ) is indeed a complex, i.e. dn−1 ◦ dn =
0.
We next aim at comparing the mapping cylinder with the mapping cone, as
defined in Definition 6.3. We consider the following morphisms of complexes
ι : X → Cyl(f ) given in degree n by ιn = (idXn , 0, 0),

0 idXn−1 0
π : Cyl(f ) → M(f ) given in degree n by πn = .
0 0 idYn
We leave the straightforward verification to the reader that these maps indeed
commute with the differentials. Clearly, the resulting sequence
ι π
0 → X −→ Cyl(f ) −→ M(f ) → 0
is a short exact sequence in C(A).
Lemma 7.21 (Mapping cylinder vs. mapping cone). Let f : X → Y be a

morphism of complexes. Then there are morphisms of complexes σ : Y →
Cyl(f ) with σn = (0, 0, idYn ) and τ : Cyl(f ) → Y with τn = (fn , 0, idYn ) such
that the following holds
(i) The following diagram with exact rows is commutative in the category
C(A) of complexes.
α(f ) β(f )
0 −→ ⏐Y −→ ⏐ ) −→ X[1]
M(f −→ 0
⏐ ⏐
⏐σ ⏐id

ι π
0 −→ X⏐ −→ ⏐ ) −→
Cyl(f M(f ) −→ 0
⏐ ⏐
⏐id ⏐τ

f
X −→ Y
(ii) τ ◦ σ = idY and σ ◦ τ is homotopic to the identity idCyl(f ) , i.e. Y and

Cyl(f ) are isomorphic in the homotopy category, and hence also in the
derived category D(A).
(iii) σ and τ are quasi-isomorphisms.
Proof. (i) It is immediately checked from the definitions (for α(f ) see Remark
6.4) that σ and τ are indeed morphisms of complexes and that all squares in
the diagram commute (in C(A), not only up to homotopy).
(ii) By definition we have τ ◦ σ = idY . On the other hand, σ ◦ τ is homotopic
to the identity via the homotopy map s = (sn ) with
⎛ ⎞
0 0 0
sn = ⎝ idXn 0 0⎠.
0 0 0
(iii) By (ii) the compositions τ ◦ σ and σ ◦ τ are homotopy equivalences,

in particular they induce the identity in homology. Thus,
Hn (σ ) ◦ Hn (τ ) = Hn (σ ◦ τ ) = Hn (idCyl(f ) ) = idHn (Cyl(f ))
and similarly Hn (τ ) ◦ Hn (σ ) = idHn (Y ) which implies that Hn (σ ) and Hn (τ ) are

isomorphisms, i.e. σ and τ are quasi-isomorphisms.
As a consequence we can now state the main result of this section.
f g
Corollary 7.22. To any short exact sequence 0 → X −→ Y −→ Z → 0 in
C(A) there exists a corresponding distinguished triangle in D(A) of the form
f g
X −→ Y −→ Z −→ X[1].
Proof. Using the notations of the previous lemma we consider the following
diagram with exact rows
ι π
0 →X
⏐ −→ Cyl(f
⏐ ) −→ M(f
⏐ ) → 0
⏐ ⏐ ⏐
⏐id ⏐τ ⏐γ

f g
0 →X −→ Y −→ Z → 0
where γ = (γn ) is defined by setting γn (x, y) := gn (y); this is easily checked
to be a morphism of complexes (using that g is a morphism of complexes
and that g ◦ f = 0 since they are consecutive maps in a short exact sequence).
Also one immediately deduces from the definitions that the above diagram
is commutative (where the left hand square already appeared in the previous
lemma).
Since id and τ (by the previous lemma) are both quasi-isomorphisms it
follows from the long exact homology sequences and the usual 5-lemma that
also γ must be a quasi-isomorphism, hence an isomorphism in the derived
category (but not necessarily in K(A), see the following remark). So we have
a morphism of triangles in D(A)
f α(f ) β(f )
X
⏐ −→ ⏐
Y −→ M(f
⏐ ) −→ X[1]
⏐
⏐ ⏐ ⏐ ⏐
⏐id ⏐id ⏐γ ⏐id

f g β(f )◦γ −1
X −→ Y −→ Z −→ X[1]
where the inverse γ −1 exists in the derived category. For the commutativity of
the second square note that by the previous lemma and the above definition of
γ we have
γ ◦ α(f ) = γ ◦ π ◦ σ = g ◦ τ ◦ σ = g ◦ id = g.
Moreover, since γ is an isomorphism in D(A) this morphism of triangles is

indeed an isomorphism of triangles, i.e. the bottom line
f g β(f )◦γ −1
X −→ Y −→ Z −→ X[1]
is isomorphic to the image of a standard triangle from K(A) under the locali-
sation functor, hence a distinguished triangle in D(A).
Remark 7.23. The crucial quasi-isomorphism γ occurring in the previous

proof is in general not a homotopy equivalence (i.e. not an isomorphism in
K(A)) and hence the isomorphism of triangles exists only in the derived cate-
gory but not yet in the homotopy category.
As an example, consider A = R-mod for a ring R and consider R-modules

X, Y as complexes concentrated in degree 0. The mapping cone of a (mod-
f
ule) morphism f : X → Y is just 0 → X −→ Y → 0, and the morphism
γ : M(f ) → Y /X is just the natural projection. If this γ is a homotopy equiv-
alence then f must be a split monomorphism. In fact, up to homotopy there
exists an inverse ρ : Y /X → Y ; for ρ ◦ γ to be homotopic to the identity on
M(f ) there must be a homotopy map s : Y → X which (when looking in
degree 1) in particular satisfies s ◦ f = idX , i.e. f splits.
This explains again the earlier example (cf. Remark 6.8) of the short exact
sequence of abelian groups
·2 ·1
0 → Z/2Z −→ Z/4Z −→ Z/2Z → 0
which does not have a corresponding distinguished triangle in K(A), because
·2
clearly the map Z/2Z −→ Z/4Z does not split.
8. Frobenius categories and stable categories

In the earlier sections we have considered categories of complexes leading to
derived categories which form an important source of examples of triangulated
categories.
In this section we shall briefly describe another source for triangulated
categories, namely stable categories of Frobenius algebras. The aim is to give
the relevant definitions and constructions and to provide some examples, in
order to prepare the ground for the later articles in this book.
For more details, in particular for a complete proof of the triangulated struc-
ture of the stable category of a Frobenius algebra we refer the reader for instance
to the well-written chapter on Frobenius categories in D. Happel’s book [5].
We start by defining an exact category, a concept introduced by D. Quillen,
which generalizes abelian categories, in the sense that an exact category has
a certain class of ‘exact triples’ as a replacement for short exact sequences
without having to be abelian itself.
Definition 8.1 (Exact category). Let A be an abelian category, and let B be

an additive subcategory of A which is full and closed under extensions (i.e. if
0 → X → Y → Z → 0 is an exact sequence in A where X and Z are objects
in B then Y is isomorphic to an object of B). Take E to be the class of all triples
X → Y → Z in B whose corresponding sequences 0 → X → Y → Z → 0
in A are exact.
Then the pair (B, E) is called an exact category.
Example 8.2.
(i) Every abelian category is an exact category; in fact, take for E the class of
all short exact sequences in A.
(ii) Let A = Ab be the category of abelian groups and B := tf-Ab the full
subcategory of torsionfree abelian groups. Then B is closed under exten-
α β
sions; in fact, let 0 → X → Y → Z → 0 be a short exact sequence with
X and Z torsionfree. Suppose ny = 0 for some y ∈ Y and n ∈ Z. Then
nβ(y) = β(ny) = 0 and hence β(y) = 0 since Z is torsionfree. By exact-
ness of the sequence it follows that y is in the image of α, say y = α(x).
But then α(nx) = nα(x) = ny = 0 which implies nx = 0 since α is a
monomorphism. From the torsionfreeness of X we deduce that x = 0 and
thus also y = 0, i.e. Y is also torsionfree. So the pair (tf-Ab, E) is an exact
category.
However, note that tf-Ab is not an abelian category (e.g. the morphism
·2
Z −→ Z does not have a cokernel in tf-Ab).
(iii) Similarly to the preceding example, consider the category t-Ab with
objects the abelian groups containing torsion elements and the trivial
group. This is a full subcategory of Ab which is closed under extensions;
in fact, if 0 → X → Y → Z → 0 is a short exact sequence of abelian
groups and X and Z contain torsion elements, then Y contains the torsion
elements of X if X is nonzero, and otherwise Y has the torsion elements
of Z. So, (t-Ab, E) is also an exact category.
Definition 8.3. An exact category (B, E) is called a Frobenius category if the

following holds:
(i) Projective and injective objects coincide (where an object I of B is called

injective if every exact triple in E of the form I → Y → Z splits; and an
object P of B is called projective if every exact triple in E of the form
X → Y → P splits).
(ii) The category has enough projective objects and enough injective objects,
i.e. for every object X in B there exist triples from E of the form X →
I → X and X → I → X where I and I are injective.
Example 8.4.
(i) A (finite-dimensional) algebra over a field K is called a Frobenius
algebra if there exists a non-degenerate associative bilinear form on .
Equivalently, if there exists a linear form π : → K such that the kernel
of π does not contain any nonzero left ideal of .
The notion of Frobenius algebra is closely related to that of selfinjec-

tive algebras (for which by definition projective and injective modules
coincide). In fact, every Frobenius algebra is selfinjective, and every basic
selfinjective algebra is Frobenius. For more details we refer to the notes by
R. Farnsteiner [3].
Then, for any Frobenius algebra the category of finite-dimensional
-modules is a Frobenius category. For instance, for a finite group G, the
category of finitely generated modules over the group algebra KG is a
Frobenius category.
(ii) Consider the abelian category B = Ab of abelian groups.
An object A in Ab is injective if and only if it is divisible (i.e. for every
·n
n ∈ Z \ {0} the multiplication map A −→ A is surjective); e.g. Q, R or
Q/Z are injective abelian groups. On the other hand, abelian groups are
nothing but modules over Z and since Z is a principal ideal domain, the
projective objects are precisely the free objects, i.e. direct sums of copies
of Z. In particular, in Ab, projective and injective objects do not coincide
and hence Ab can not be a Frobenius category.
Definition 8.5. Let (B, E) be a Frobenius category. For objects X and Y in B

let Inj(X, Y ) denote those morphisms from X to Y which factor through some
injective object.
The stable category B of the Frobenius category (B, E) has the same objects
as B; the morphisms are equivalence classes of morphisms modulo those
factoring through injective objects, i.e.
HomB (X, Y ) := Hom(X, Y ) = HomB (X, Y )/ Inj(X, Y ).
Example 8.6 (Dual numbers). Let K be a field and consider the algebra
= K[X]/(X 2 ). This 2-dimensional algebra is a Frobenius algebra; in fact,
π(a + bX) := b defines a linear form on such that its kernel does not contain
any nonzero left ideal.
The category of finitely generated -modules is then a Frobenius category. It
has only two indecomposable objects, itself and the one-dimensional simple
module K, where is the only injective (and projective) module.
In the corresponding stable module category several module homomor-
phisms vanish, e.g. we have that Hom(, ) = 0, Hom(, K) = 0 and
Hom(K, ) = 0; on the other hand Hom(K, K) remains 1-dimensional since
the isomorphism can not factor through the injective module .
Our main aim is to describe how the stable category of a Frobenius category
(B, E) carries the structure of a triangulated category. To this end we shall briefly
describe the construction of a suspension functor and then of the distinguished

triangles.
ιX πX
For every object X in B we choose an exact triple X −→ I (X) −→ X
with entries from B where I (X) is an injective object, i.e. in the ambient abelian
category A there is a short exact sequence 0 → X → I (X) → X → 0. Thus,
on objects of B (and hence also on objects of B) we have a map X
→ X and
this will be the candidate for the suspension functor on objects.
For defining on morphisms, let u : X → Y be a morphism and con-
ιX πX ιY πY
sider the chosen exact triples X −→ I (X) −→ X and Y −→ I (Y ) −→ Y .
Since I (Y ) is injective there exists a morphism I (u) : I (X) → I (Y ) such
that I (u)ιX = ιY u. By considering the corresponding short exact sequences
in the ambient abelian category the morphism I (u) induces a morphism
(u) : X → Y satisfying (u)πX = πY I (u).
Lemma 8.7. The morphisms (u) are well-defined in the stable category B,
i.e. they are independent of the choice of the lifting morphisms I (u) and of the
representatives of the morphisms u.
Proof. For a morphism u the morphism (u) has been defined above as indi-
cated in the following diagram
ιX- πX-
X I (X) X
u I (u) (u)
? ? ?
ιY - πY-
Y I (Y ) Y
Now let I (u) and I˜(u) be two liftings for u, with corresponding induced mor-
˜
phisms (u) and (u) from X to Y . Then
(I (u) − I˜(u))ιX = I (u)ιX − I˜(u)ιX = ιY u − ιY u = 0.
By exactness of the top row there exists a morphism σ : X → I (Y ) such that

σ πX = I (u) − I˜(u). This implies that
((u) − (u))π
˜ X = πY I (u) − πY I (u) = πY σ πX .
˜
Since πX is an epimorphism we can deduce that (u) − (u) ˜ = πY σ . Hence

(u) − (u)
˜ factors through the injective object I (Y ), i.e. (u) = (u)
˜ in the
stable category B.
A similar argument shows that if u factors through an injective object then
also (u) factors through an injective object, i.e. (u) is independent of the
representative of the morphism u in B. We leave the details to the reader.
Our above construction of the objects X and of the morphisms (u) used
a fixed choice of exact triples X → I (X) → X. The following lemma shows
that this construction does not depend on the choice of the exact triples, more
precisely, a different choice leads to naturally isomorphic functors. In particular,
the object X is uniquely defined up to isomorphism in the stable category B.
Lemma 8.8. For any object X let X → I (X) → X and X → I (X) → X

be exact triples in B where I (X) and I (X) are injective. Then X and X
are isomorphic in the stable category B. Moreover, there is a natural transfor-
mation β : → such that each βX : X → X is an isomorphism, i.e.
the functors and are isomorphic.
Proof. In the ambient abelian category A we have short exact sequences
ιX πX ιX πX
0 → X −→ I (X) −→ X → 0 and 0 → X −→ I (X) −→ X → 0.
Since I (X) and I (X) are injective in B there are morphisms αX : I (X) →
I (X) and αX : I (X) → I (X) making the left hand squares in the following
diagram commutative; moreover since the rows are short exact sequences these
morphisms induce morphisms βX and βX also making the right hand squares
commutative.
0 -X ιX-
I (X)
πX-
X -0
αX βX
? ?
-X ιX- πX - -0

0 I (X) X

αX
βX
? ?
0 -X ιX-
I (X)
πX-
X -0
By commutativity it follows that
(αX αX − idI (X) )ιX = αX ιX − ιX = ιX − ιX = 0.
Using exactness of the top row implies the existence of a morphism σX :

X → I (X) such that σX πX = αX αX − idI (X) . Again using commutativity of
the above diagram we then have that
πX σX πX = πX αX αX − πX = (βX βX − idX )πX .

Since πX is an epimorphism we deduce πX σX = βX βX − idX , i.e. βX βX −

idX factors through the injective object I (X) and thus βX βX = idX in the
stable category B.
An analogous argument shows that also βX βX = id X in B. Thus, X and

X are isomorphic in B, as claimed.
For naturality, consider different choices of exact triples for objects X and
Y and a morphism u : X → Y , as in the following commutative diagrams
ιX- πX- ιX- πX -

X I (X) X X I (X) X
u I (u) (u) u I (u) (u)

? ? ? ? ? ?
ιY - πY- ιY - πY -
Y I (Y ) Y Y I (Y ) Y
We claim that β induces a natural isomorphism between the functors and

resulting form different choices of exact triples. From the first part of the
proof we already know that each βX is an isomorphism in B. It remains to
show that we indeed have a natural transformation, i.e. we have to show that
βY (u) = (u)βX in the stable category.
Note that by commutativity of the above diagrams we have
(αY I (u) − I (u)αX )ιX = αY ιY u − I (u)ιX = ιY u − I (u)ιX = 0.
Hence by exactness there exists a morphism τ : X → I (Y ) such that τ πX =

αY I (u) − I (u)αX . It then follows that
πY τ πX = πY (αY I (u) − I (u)αX ) = βY πY I (u) − (u)πX αX

= βY (u)πX − (u)βX πX = (βY (u) − (u)βX )πX .
Since πX is an epimorphism this implies that πY τ = βY (u) − (u)βX , i.e.

βY (u) − (u)βX factors through the injective object I (Y ) and hence we
have βY (u) = (u)βX in the stable category B.
Hence we have a well-defined functor on the stable category B. This can

be shown to be an autoequivalence. Under certain assumptions it is even an
automorphism; for details on this subtle issue see Happel’s book [5, Section
I.2].
We now describe the construction of distinguished triangles in the stable
category B. Let X, Y be objects in B and let u : X → Y be a morphism. For X
ι π
we have from the above construction an exact triple X → I (X) → X where
I (X) is injective.
In the additive category B there exists a coproduct I ⊕ Y , with morphisms

ιI : I → I ⊕ Y and ιY : Y → I ⊕ Y satisfying the universal property given
in Remark 1.2. Now we form the pushout of the morphisms ι : X → I and
u : X → Y . More precisely, the pushout M(u) is defined as the cokernel of
the morphism ιI ι − ιY u : X → I ⊕ Y . By definition, this cokernel is an object
M(u) together with a morphism c : I ⊕ Y → M(u) such that c(ιI ι − ιY u) =
0 and satisfying the universal property for cokernels given in Section 2. In
particular, the left hand square in the following diagram is commutative
ι π
X
⏐ −→ I⏐(X) −→ X
⏐
⏐ ⏐ ⏐
⏐u ⏐cιI ⏐id

cιY
Y −→ M(u) X
We wish to complete this diagram with a morphism w : M(u) → X. To this

end recall that from the properties of a coproduct there is a unique morphism f :
I ⊕ Y → X such that f ιI = π and f ιY = 0. Using this morphism f in the
universal property for the cokernel M(u) we deduce the existence of a (unique)
morphism w : M(u) → X such that wc = f . It follows that wcιI = f ιI =
π, i.e. the following diagram of objects and morphisms in B is commutative
ι π
X
⏐ −→ I⏐(X) −→ X
⏐
⏐ ⏐ ⏐
⏐u ⏐cιI ⏐id

cιY w
Y −→ M(u) −→ X
Note that since B is closed under extensions the cokernel will again be an object
in B (and not only in the ambient abelian category A).
The images in the stable category B of any triangles of the form
u cιY w
X −→ Y −→ M(u) −→ X
are called standard triangles in B. As usual, the set of distinguished triangles

in B is formed by the set of all triangles in B which are isomorphic (in B !) to
a standard triangle.
The main structural result on stable categories of Frobenius categories is
then the following; for a detailed proof we refer to Happel’s book [5, Chapter
I.2].
Theorem 8.9. Let (B, E) be a Frobenius category. With the above suspension
functor and the collection of distinguished triangles just defined, the stable
category B is a triangulated category.
Remark 8.10. A triangulated category is called algebraic (in the sense of

B. Keller, see [9]) if it is equivalent as a triangulated category to the stable
category of a Frobenius category. For more details on algebraic and non-
algebraic triangulated categories we refer to S. Schwede’s article in this volume
[14], see also [15]. Strikingly, there are triangulated categories which are neither
algebraic nor topological [11].
References
[1] P. Balmer, Triangulated categories with several triangulations. Available
from the author’s web page at

∼

[2] A. Dold, D. Puppe, Homologie nicht-additiver Funktoren. Ann. Inst. Fourier
Grenoble 11 (1961) 201–312.
[3] R. Farnsteiner, Self-injective algebras: I. The Nakayama permutation, II. Compar-
ison with Frobenius algebras, III. Examples and Morita equivalence, IV. Frobe-
nius algebras and coalgebras. Available at

∼
[4] S.I. Gelfand, Y.I.Manin, Methods of Homological Algebra. Second edition.
Springer Monographs in Mathematics. Springer-Verlag, Berlin, 2003.
[5] D. Happel, Triangulated categories in the representation theory of finite-
dimensional algebras. London Mathematical Society Lecture Note Series, 119.
Cambridge University Press, Cambridge, 1988.
[6] A. Hubery, Notes on the octahedral axiom. Available from the author’s web page
at
∼

[7] M. Kashiwara, P. Schapira, Sheaves on manifolds. Grundlehren der Mathematis-
chen Wissenschaften, 292. Springer-Verlag, Berlin, 1990.
[8] M. Kashiwara, P. Schapira, Categories and sheaves. Grundlehren der Mathema-
tischen Wissenschaften 332. Springer-Verlag, 2006.
[9] B. Keller, On differential graded categories. International Congress of Mathemati-
cians. Vol. II, 151–190, Eur. Math. Soc., Zürich, 2006.
[10] H. Krause, Localization theory for triangulated categories. This volume.
[11] F. Muro, S. Schwede, N. Strickland, Triangulated categories without models.
Invent. Math. 170 (2007), 231-241.
[12] A. Neeman, Triangulated categories. Annals of Mathematics Studies, 148. Prince-
ton University Press, Princeton, NJ, 2001.
[13] A. Neeman, New axioms for triangulated categories. J. Algebra 139 (1991), 221–
255.
[14] S. Schwede, Algebraic versus topological triangulated categories. This volume.
[15] S. Schwede, Torsion invariants for triangulated categories. Preprint (2009), avail-
able from the author’s web page at http://www.math.uni-bonn.de/∼schwede.
[16] J.-L. Verdier, Des catégories dérivées des catégories abéliennes. Astérisque No.
239 (1996).
[17] C.A. Weibel, An introduction to homological algebra. Cambridge Studies in

Advanced Mathematics, 38. Cambridge University Press, Cambridge, 1994.
Thorsten Holm
Leibniz Universität Hannover, Institut für Algebra, Zahlentheorie und Diskrete
Mathematik, Welfengarten 1, D-30167 Hannover, Germany
E-mail address:

URL:

Peter Jørgensen
School of Mathematics and Statistics, Newcastle University, Newcastle upon Tyne
NE1 7RU, United Kingdom
E-mail address:

URL:

Cohomology over complete intersections
via exterior algebras
luchezar l. avramov and
srikanth b. iyengar
To Karin Erdmann on her 60th birthday.
Abstract. A general method for establishing results over a commutative com-

plete intersection local ring by passing to differential graded modules over a
graded exterior algebra is described. It is used to deduce, in a uniform way,
results on the growth of resolutions of complexes over such local rings.
Introduction
This paper concerns homological invariants of modules and complexes over
complete intersection local rings. The goal is to explain a method by which
one can establish in a uniform way results over such rings by deducing them
from results on DG (that is, differential graded) modules over a graded exterior
algebra, which are often easier to prove. A secondary purpose is to demonstrate
the use of numerical invariants of objects in derived categories, called ‘levels’,
introduced in earlier joint work with Buchweitz and Miller [5]; see Section 1.
Levels allow one to track homological and structural information under changes
of rings or DG algebras, such as those involved when passing from complete
intersections to exterior algebras.
We focus on the complexity and the injective complexity of a complex M
over a complete intersection ring R. These numbers measure, on a polynomial
scale, the rate of growth of the minimal free resolution and the minimal injective
resolution of M, respectively. The relevant basic properties are established in
Section 2.
Complexities can be expressed as dimensions of certain algebraic varieties,
attached to M in [2], and earlier proofs of key results relied on that theory. In
Date: April 28, 2010.

2000 Mathematics Subject Classification. 13D02, 16E45; 13D07, 13D25, 13H10, 20J06.
Key words and phrases. Complexity, exterior algebra, complete intersection local ring,
differential graded Hopf algebra, Bernstein-Gelfand-Gelfand correspondence.
Research partly supported by NSF grants DMS 0803082 (L.L.A) and DMS 0602498 (S.B.I). Part
of the work of the second author was done at the University of Paderborn on a visit supported by
a Forschungspreis from the Humboldtstiftung.
52
Cohomology over complete intersections 53
Section 6 we deduce them from results on differential graded Hopf algebras,

presented with complete proofs in Section 4. These pleasingly simple proofs
build on nothing more than basic homological algebra, summarized in Section 3.
In the last three sections of this article the goal is to link the complexity
of M and the Loewy length of its homology modules. In [5] such a result is
deduced from a more general statement, which applies to arbitrary local rings.
From that paper we import DG versions of the New Intersection Theorem,
recalled in Section 7, and of the Bernstein-Gelfand-Gelfand correspondence,
which we refine in Section 8. This allows us to establish a result on complexities
over exterior algebras, which we then translate in Section 9 into the desired
link between Loewy length and complexity over complete intersection local
rings.
1. Levels
In this section A denotes a DG algebra.
We write D(A) for the derived category of DG A-modules; it is a triangulated
category with shift functor denoted Σ. The underlying graded object of a DG
object X is denoted X . Rings are treated as DG algebras concentrated in
degree zero; over a ring DG modules are simply complexes, and modules are
DG modules concentrated in degree zero; see [5, §3] for more details and
references. Unless specified otherwise, all DG modules have left actions.
A non-empty subcategory C of D(A) is said to be thick if it is an additive
subcategory closed under retracts, and every exact triangle in D(A) with two
vertices in C has its third vertex in C; thick subcategories are triangulated.
Given a DG A-module X, we write thickA (X) for the smallest thick sub-
category containing X. The existence of such a subcategory can be seen by
realizing it as the intersection of all thick subcategories of D(A) that contain X.
Alternatively, the objects of thickA (X) can be built from X in a series of steps,
described below.
1.1. For every X in D(A) and each n ≥ 0 we define a full subcategory thicknA (X)
of thickA (X), called the nth thickening of X in A, as follows. Set thick0A (X) =
{0}; the objects of thick1A (X) are the retracts of finite direct sums of shifts of X.
For each n ≥ 2, the objects of thicknA (X) are retracts of those U ∈ D(A) that
appear in some exact triangle U → U → U → ΣU with U ∈ thickn−1 A (X)
and U ∈ thick1A (X).
Every thickening of X is clearly embedded in the next one; it is also clear
that their union is a thick subcategory of A containing X, which is therefore
54 L. L. Avramov and S. B. Iyengar
equal to thickA (X). Thus, thickA (X) is equipped with a filtration

{0} = thick0A (X) ⊆ thick1A (X) ⊆ · · · ⊆ thicknA (X) = thickA (X) .
n∈N
To each object U in D(A) we associate the number
A (U ) = inf{n ∈ N | U ∈ thickA (X)}

levelX n
and call it the X-level of U in D(A). It measures the number of extensions needed
to build U out of X. Evidently, levelX A (U ) < ∞ is equivalent to U ∈ thickA (X).
We refer the reader to [5, §§2–6] for a systematic study of levels. The
properties used explicitly in this paper are recorded below.
1.2. Let U , X, and Y be objects in D(A).
(1) If thick1A (X) = thick1A (Y ), then one has thicknA (X) = thicknA (Y ) for all n,
and hence an equality levelX A (U ) = levelA (U ).
Y
(2) If B is a DG algebra and j : D(A) → D(B) is an exact functor, then

j(X)
A (U ) ≥ levelB (j(U )) .
levelX
Equality holds when j is an equivalence.
A level of interest in this paper is related to the notion of Loewy length.
1.3. Let B be a ring and k a simple B-module. For each B-module H , kB H
denotes the smallest integer l ≥ 0 such that H has a filtration by B-submodules
0 = H 0 ⊆ H 1 ⊆ · · · ⊆ H l−1 ⊆ H l = H
with every H i /H i−1 isomorphic to a sum of copies of k; if no such filtration

exists, we set kB H = ∞. If B has a unique maximal left ideal m, then k is
isomorphic to B/m and kB H is equal to the Loewy length of H , defined by
the formula
B H = inf{l ∈ N | ml H = 0} ;
see [5, 6.1.3]. If lengthB H is finite, then so is B H , and the converse holds
when H is noetherian; see [5, 6.2(4)].
We say that the DG algebra A is non-negative if Ai = 0 for i < 0. When

this is the case, there is a canonical morphism of DG algebras A → H0 (A),
called the augmentation of A; it turns every H0 (A)-module into a DG
A-module.
1.4. When A is non-negative and k is a simple H0 (A)-module one has:

(1) The number levelkA (U ) is finite if and only if the H0 (A)-module

n∈Z Hn (U ) admits a finite filtration with subfactors isomorphic to k.
(2) When levelkA (U ) is finite the following inequality holds:

levelkA (U ) ≤ kH0 (A) Hn (U ) .
n∈Z
(3) When H0 (A) = k and the k-module H(U ) is finite, one has an inequality
levelkA (U ) ≤ card{n ∈ Z | Hn (U ) = 0} .
Indeed, (1) and (2) are contained in [5, 6.2(3)], while (3) is extracted from
[5, 6.4].
2. Complexities
In this section we introduce a notion of complexity for DG modules over a
suitable class of DG algebras, and establish some of its elementary properties.
We begin with a reminder of the construction of derived functors on the derived
category of DG modules over a DG algebra; see [6, §1] for details.
2.1. Let A be a DG algebra and Ao its opposite DG algebra.
A semifree filtration of a DG A-module F is a filtration
0 = F 0 ⊆ F 1 ⊆ · · · ⊆ F n−1 ⊆ F n ⊆ · · ·

by DG submodules with n0 F n = F and each DG module F n /F n−1 iso-
morphic to a direct sum of suspensions of A; when one exists F is said to be
semifree. When F is semifree its underlying graded A -module F is free, and
the functors HomA (F, −) and (−⊗A F ) preserve quasi-isomorphisms of DG
A-modules.
A semifree resolution of a DG module U is a quasi-isomorphism F → U
of DG modules with F semifree; such a resolution always exists. If U → V is
a quasi-isomorphism of DG A-modules and G → V is a semifree resolution,
then there is a unique up to homotopy morphism F → G of DG A-modules
such that the composed maps F → G → V and F → U → V are homotopic.
These properties imply that after choosing a resolution FU → U for each
U , the assignments (U, V )
→ HomA (FU , V ) and (W, U )
→ W ⊗A FU define
exact functors
RHomA (−, −) : D(A)o × D(A) → D(Z) and
− ⊗LA − : D(A ) × D(A) → D(Z) ,
o
respectively. In homology, they define graded abelian groups
ExtA (U, V ) = H(RHomA (U, V )) and TorA (W, U ) = H(W ⊗LA U ) .
In case of modules over rings these are the classical objects.
We want to measure, on a polynomial scale, how the ‘size’ of ExtnA (U, V )

grows when n goes to infinity. In order to do this we use the notion of complex-
ity of a sequence (bn )n∈Z of non-negative numbers, defined by the following
equality

there is a number a ∈ R such that

cx (bn ) := inf d ∈ N .
bn ≤ and−1 holds for all n 0
Throughout the rest of this section, (A, m, k) denotes a local DG algebra,

by which we mean that A is non-negative, A0 is a noetherian ring contained in
the center of A, m is the unique maximal ideal of A0 , and k is the field A0 /m.
The complexity of a pair (U, V ) of DG A-modules is the number

cxA (U, V ) = cx rankk (ExtnA (U, V ) ⊗A0 k) .
The complexity and the injective complexity of U are defined, respectively, by
cxA U = cxA (U, k) and inj cxA U = cxA (k, U ) .
When A is a local ring cxA U is the polynomial rate of growth of rankA Fn ,

where F is a minimal free resolution of U , while inj cxA U is that of the
multiplicity of an injective envelope of k in I n , where I is a minimal injective
resolution of U .
The following properties of these invariants have been observed before for
modules, and for complexes, over local rings. We extend them to handle DG
modules over DG algebras, sometimes with alternative proofs.
Lemma 2.2. If U is in thickA (X) for some DG A-module X, then one has
cxA U ≤ cxA X and inj cxA U ≤ inj cxA X .
Equalities hold in case thickA (U ) = thickA (X).
Proof. We verify the inequality for complexities; a symmetric argument yields

the one for injective complexities.
The number h = levelX A (U ) is finite; we induce on it. When h equals one U

is a retract of sj =1 Σij X for some integers i1 , . . . , is , so ExtnA (U, k) is a direct
n+i
summand of sj =1 ExtA j (X, k), and the desired inequality is clear. For h ≥ 1
one has an exact triangle U → U → U → ΣU with levelX
A (U ) ≤ h − 1

A (U ) = 1; the associated cohomology exact sequence shows that the
and levelX
induction hypothesis implies the desired inequality for complexities.
Lemma 2.3. If F is a finite free complex of A0 -modules with H(F ) = 0, then
cxA (U ⊗A0 F ) = cxA U and inj cxA (U ⊗A0 F ) = inj cxA U .
Proof. Since F is a finite free complex, in D(A) one has isomorphisms
RHomA (k, U ⊗A0 F ) RHomA (k, U ) ⊗LA0 F RHomA (k, U ) ⊗Lk (k⊗A0 F ) .
This yields, for each integer n, an isomorphism of k-vectorspaces

ExtnA (k, U ⊗A0 F ) ∼
= A (k, U ) ⊗k Hi (k ⊗A0 F ) .
Extn+i
i∈Z
Since Hi (k ⊗A0 F ) is finite for each i, and zero for |i| 0 but not for all i,
the equality of injective complexities follows. The argument for complexities
is similar.
2.4. For a morphism of DG algebras ϕ : A → B, let ϕ∗ be the functor forgetting

the action of B. One then has an adjoint pair of exact functors
(B⊗LA −)
/
D(A) o D(B)
ϕ∗
that are inverse equivalences if ϕ is a quasi-isomorphism; see [5, 3.3.1, 3.3.2].

If, furthermore, U is a DG A-module, V is a DG B-module, and μ : U → V
is a ϕ-equivariant quasi-isomorphism, then by [5, 3.3.3] one has canonical
isomorphisms
U ϕ∗ (V ) and B ⊗LA U V .
A DG algebra B is said to be quasi-isomorphic to A if there exists a

chain
/ A0 o / ··· o / Ai o
A A1 Ai−1 B
of quasi-isomorphisms of DG algebras. Such a chain induces a unique equiva-

lence
j : D(A) → D(B) ,
of triangulated categories; if Bi = 0 for i < 0, then [5, 3.6] yields
j(H0 (A)) H0 (B) .

Lemma 2.5. If (B, n, l) is a local DG algebra quasi-isomorphic to A, then

each equivalence of derived categories j : D(A) → D(B) as in 2.4 induces an
isomorphism
j(k) l
in D(B), and for all U and V in D(A) there is an equality
cxA (U, V ) = cxB (j(U ), j(V )) .
Proof. As H0 (A) is equal to A0 /∂(A1 ), it is a local ring with maximal ideal

m/∂(A1 ). Similarly, H0 (B) = B0 /∂(B1 ) is a local ring with maximal ideal
n/∂(B1 ). The isomorphism of 2.4 maps j(m/∂(A1 )) to n/∂(B1 ), and so induces
j(k) l.
Since j is an equivalence, in D(B) there is an isomorphism
j(RHomA (U, V )) RHomB (jU, jV ) .
Passing to homology one obtains an isomorphism of graded modules
ExtA (U, V ) ∼
= ExtB (jU, jV )
that is equivariant over the isomorphism H(A) ∼= H(B) of graded algebras

induced by the chain of quasi-isomorphisms inducing j. This yields an isomor-
phism
ExtA (U, V ) ⊗H0 (A) k ∼

= ExtB (jU, jV ) ⊗H0 (B) l
of graded vector spaces that is equivariant over the isomorphism k ∼

= l.
Sometimes there exist an alternative way for computing complexity:
Lemma 2.6. If the ring A0 is artinian, then there is an equality
cxA (U, V ) = cx (lengthA0 ExtnA (U, V )) .
Proof. Set rn = rankk (ExtnA (U, V ) ⊗A0 k). By hypothesis ms = 0 holds for
some integer s, so by Nakayama’s Lemma the A0 -module ExtnA (U, V ) is min-
imally generated by rn elements. Thus, one has surjective homomorphisms
k rn ← ExtnA (U, V ) ← Ar0n
of A0 -modules. They yield inequalities
rn ≤ lengthA0 ExtnA (U, V ) ≤ lengthA0 (A0 )rn
that evidently imply cx (rn ) = cx (lengthA0 ExtnA (U, V )).

3. Composition products
Let A denote a DG algebra, and let U , V , and W be DG A-modules. In this
section we recall the construction of products in cohomology of DG A-modules.
Construction 3.1. For all integers i, j there exist composition products

j i+j
ExtA (V , W ) × ExtiA (U, V ) → ExtA (U, W ) .
Indeed, choose a semifree resolution F → U ; thus ExtiA (U, V ) =

Hi (HomA (F, V )). An element [α] ∈ ExtiA (U, V ) is then the homotopy class
of a degree −i chain map α : F → V of DG modules. Similarly, let [β] ∈
j
ExtA (V , W ) be the class of a chain map β : G → W of degree −j , where
G → V is a semifree resolution. As F is semifree and F → U is a quasi-
isomorphism there is a unique up to homotopy chain map α : F → G whose
composition with G → V is equal to α. One defines the product [β][α] in
i+j
ExtA (U, W ) to be the class of β ◦ α: F → W.
It follows from their construction that composition products are associative
in the obvious sense. In particular, ExtA (U, U ) and ExtA (V , V ) are graded DG
algebras, and that ExtA (U, V ) is a graded left-right ExtA (V , V )-ExtA (U, U )-
bimodule.
Let A and B be DG algebras over a commutative ring k. Their tensor product

is a DG algebra, with underlying complex A ⊗k B and multiplication defined
by

(a ⊗ b) · (a ⊗ b ) = (−1)|b||a | (aa ⊗ bb ) ,
where |a| denotes the degree of a. For each DG B-module X the complexes
U ⊗k X and Homk (X, U ) have canonical structures of DG module over A ⊗k
B, given by
(a ⊗ b) · (u ⊗ x) = (−1)|b||u| (au ⊗ bx) ;

(a ⊗ b)(γ ) (x) = (−1)|b||γ | aγ (bx) .
Construction 3.2. Let k be a field, and let A and B be DG algebras with
Ai = 0 = Bi for i < 0 and A0 = k = B 0 .
Let F → k and G → k be semifree resolutions over A and B, respectively, and
ω : HomA (F, k) ⊗k HomB (G, k) → HomA⊗k B (F ⊗k G, k)

be the morphism given by ω(α ⊗ β) (a ⊗ b) = (−1)|β||a| α(a)β(b). Compos-
ing H(ω) with the Künneth isomorphism [α] ⊗k [β]
→ [α ⊗ β] one obtains a
homomorphism
H(HomA (F, k)) ⊗k H(HomB (G, k)) → H(HomA⊗k B (F ⊗k G, k))
The map F ⊗k G → k ⊗k k = k evidently is a semifree resolution over A ⊗k
B, so the preceding homomorphisms defines a Künneth homomorphism
κ : ExtA (k, k) ⊗k ExtB (k, k) → ExtA⊗k B (k, k) .
It is follows from the definition that this is a homomorphism of k-algebras.
Lemma 3.3. If rankk Ai is finite for each i, then κ is an isomorphism.
Proof. Under the hypothesis on A, one can choose an F that has a semifree
filtration where each free graded A -module (F n /F n−1 ) of finite rank; for
instance, take F to be the classical bar-construction, see [12]. In this case the
map ω from Construction 3.2 is bijective, so H(ω) is an isomorphism.
4. Differential Graded Hopf algebras

In this section k is a field and A is a DG Hopf algebra over k; that is, A is a
non-negative DG algebra with A0 = k, and with a morphism : A → A ⊗k A
of DG k-algebras, called the comultiplication, satisfying (εA ⊗ A) = idA =
(A ⊗k εA ), where εA : A → k is the canonical surjection.
The principal results in this section, Theorems 4.7 and 4.9, establish sym-
metry properties of complexities. The arguments in the proof exploit operations
on DG modules provided by the Hopf algebra structure on A.
4.1. When A is a DG Hopf algebra the DG (A ⊗k A)-modules U ⊗k V and
Homk (U, V ) acquire structures of DG A-modules through . The isomor-
phisms
U ⊗k k ∼ =U ∼= k ⊗k U (4.1.1)
∼
Homk (k, V ) = V (4.1.2)
HomA (U ⊗k V , W ) ∼
= HomA (U, Homk (V , W )) (4.1.3)
are compatible with the respective DG A-structures.
Lemma 4.2. Let F → k be a semifree resolution and G a semifree DG
module. The induced morphism π : F ⊗k G → k ⊗k G ∼
= G is a homotopy
equivalence.
Proof. The morphism π is a quasi-isomorphism because k is a field. As
G is semifree, the identity map on G lifts through π to give a morphism
κ : G → F ⊗k G such that π κ is homotopic to idG ; see 2.1. Since F is semifree

it follows from (4.1.3) that the functor HomA (F ⊗k G, −) preserves quasi-
isomorphisms. Noting that π κπ is homotopic to π idF ⊗k G , one thus gets that
κπ is homotopic to idF ⊗k G .
Construction 4.3. Let A be a DG Hopf algebra over k and set S = ExtA (k, k).
Let V be a DG A-module. Choose semifree resolutions F → k and G → V
over A. The assignment ψ
→ ψ ⊗k G defines a morphism
HomA (F, F ) → HomA (F ⊗k G, F ⊗k G)
of DG algebras. Since F ⊗k G is homotopy equivalent to G, by Lemma 4.2, in

homology it induces a homomorphism of graded k-algebras
ζV : S → ExtA (V , V ) .
The graded center of a graded algebra B consists of the elements c in B that

satisfy bc = (−1)|b||c| cb for all b ∈ B. When every c ∈ B has this property B
is said to be graded-commutative; every graded right B-module X then has a
canonical structure of left B-module, defined by setting bx = (−1)|b||x| xb.
Proposition 4.4. Let A be a DG Hopf k-algebra, and let U, V be DG A-

modules.
The algebra S = ExtA (k, k) is then graded-commutative and ζV (S) is con-
tained in the graded center of ExtA (V , V ). Moreover, the S-module structures
on ExtA (U, V ) induced via ζV and ζU coincide.
Proof. We verify the second assertion; the first one follows, as ζk = idS .
We may assume U and V are semifree DG A-modules. Let F → k be a
semifree resolution. By Lemma 4.2, the homology of the complex HomA (F ⊗k
U, F ⊗k V ) is ExtA (U, V ). It implies also that any chain map F ⊗k U →
F ⊗k V is homotopy equivalent to one of the form F ⊗k μ. For any chain map
σ : F → F the compositions
σ ⊗U F ⊗μ
F ⊗k U −−→ F ⊗k U −−→ F ⊗k V
F ⊗μ σ ⊗V
F ⊗k U −−→ F ⊗k V −−→ F ⊗k V
are σ ⊗ μ and (−1)|μ||σ | σ ⊗ μ, respectively. Hence the S-actions on

ExtA (U, V ) induced via ζU and ζV coincide up to the usual sign. This is
the desired result.
Proposition 4.5. Let A be a DG Hopf k-algebra, and U and V be DG A-

modules.
If the k-algebra S = ExtA (k, k) is finitely generated, and the k-vector spaces
H(U ) and H(V ) are finite, then ExtA (U, V ) is finite over S, ExtA (U, U ), and
ExtA (V , V ).
Proof. Since S acts on ExtA (U, V ) through ExtA (U, U ), finiteness over
the former implies finiteness over the latter. The same reasoning applies
with ExtA (U, U ) replaced by ExtA (V , V ), so it suffices to prove finiteness
over S.
We claim that the following full subcategory of D(A) is thick:
C = {W ∈ D(A) |the S-module ExtA (W, k) is noetherian} .
Indeed, it is clear that C is closed under retracts and shifts; furthermore, every
exact triangle W → W → W → ΣW in D(A) induces an exact sequence
ExtA (W , k) → ExtA (W, k) → ExtA (W , k)
of graded S-modules, so if W and W are in C, then so is W .

The condition rankk H(U ) < ∞ implies U ∈ thickA (k); see 1.4(1). The
finitely generated k-algebra S is graded-commutative by Proposition 4.4, and
thus noetherian; this implies k ∈ C. Now from the thickness of C we conclude
U ∈ C.
From a similar argument, now considering the subcategory
{W ∈ D(A) | the S-module ExtA (U, W ) is noetherian} ,
one deduces that the S-module ExtA (U, V ) is finitely generated.
The preceding result allows one to draw conclusions about complexities

from classical results in commutative algebra. This may not be immediately
clear to casual users of the subject, as standard textbooks leave out important
parts of the story by focusing early on on ‘standard graded’ algebras. In the
following discussion we refer to Smoke’s very readable and self-contained
exposition in [15].
4.6. Let S be a graded-commutative k-algebra, generated over k by finitely

many elements of positive degrees, and C a finitely graded S-module. Replacing
generators by their squares, one sees that C is also finite over a finitely generated
commutative k-algebra S , so there exists a Laurent polynomial pC (t) ∈ Z[t ±1 ],
such that

c
rankk (Cn )t n = pC (t) (1 − t dj ) ;
n∈Z j =1
this is the Hilbert-Serre Theorem, see [15, 4.2] or [13, 13.2]. By [15, 5.5],
the order of the pole at t = 1 of the rational function above is equal to the
Krull dimension of C; that is, to the supremum of the lengths of chains of
homogeneous prime ideals in S containing the annihilator of C. The order
of the pole is independent of the choice of S , so one gets a well defined
notion of Krull dimension of C over S; let dimS C denote this number, and
set dim S = dimS S. By decomposing the rational function above into prime
fractions, see [3, §2] for details, one easily obtains
dimS C = cx ( rankk (Cn )) .
Theorem 4.7. If A is a DG Hopf k-algebra with S = ExtA (k, k) finitely gen-
erated as k-algebra, and U a DG A-module with H(U ) finite over k, then one
has
inj cxA U = cxA (U, U ) = cxA U ≤ cxA k = dim S .
If, furthermore, V is a DG A-module with H(V ) finite over k, then one has
dimS (ExtA (U, V )) = cxA (U, V ) ≤ min{cxA (U, U ), cxA (V , V )} .
Proof. The expression for cxA (U, V ) comes from Propositions 4.4, 4.5, and
4.6. By Proposition 4.5, ExtA (U, V ) is a finite module over ExtA (U, U ) and
ExtA (V , V ), whence the upper bounds on cxA (U, V ). For (k, U ) and (U, k)
they yield
inj cxA U ≤ cxA (U, U ) ≥ cxA U ≤ cxA k .
To prove cxA (U, U ) ≤ cxA U we show that the full subcategory
D = {W ∈ D(A) | cxA (U, W ) ≤ cxA U }
of D(A) contains U . One evidently has k ∈ D, and 1.4(1) gives U ∈ thickA (k),
so it suffices to prove that D is thick. Closure under direct summands and shifts
is clear. An exact triangle W → W → W → ΣW in D(A) yields an exact
sequence
ExtnA (U, W ) → ExtnA (U, W ) → ExtnA (U, W )
for every n ∈ Z. They imply inequalities

rankk ExtnA (U, W ) ≤ rankk ExtnA (U, W ) + rankk ExtnA (U, W ) .
Thus, if W and W are in D, then, Lemma 2.6 gives
cxA (U, W ) ≤ max{cxA (U, W ), cxA (U, W )} ≤ cxA U .
A similar argument, now using ExtA (−, U ), yields inj cxA U = cxA (U, U ).
The equality cxA k = dim S comes from the Hilbert-Serre theorem,

see 4.6.
To compare cxA (U, V ) and cxA (V , U ) we need one more lemma.
Lemma 4.8. Let A be a DG Hopf k-algebra. For all U, V ∈ D(A) the assign-
ment (u ⊗ v)
→ [α
→ (−1)|v||α| uα(v)] defines a morphism
U ⊗k V → Homk (V ∗ , U )
in D(A); it is an isomorphism when rankk H(V ) is finite.
Proof. It is a routine calculation to verify that the map in question is com-
patible with the DG A-module structures. It thus defines for each X ∈ D(A)
a natural morphism ηX : U ⊗k X → Homk (X∗ , U ). It is easy to see that those
X for which ηX is an isomorphism form a thick subcategory of D(A). It con-
tains k and hence contains every DG A-module V with rankk H(V ) finite;
see 1.4(1).
The next result extends the equality cxA U = inj cxA U from Theorem 4.7.
Theorem 4.9. If A is a DG Hopf k-algebra with ExtA (k, k) finitely generated
as k-algebra, and U and V are DG A-modules with H(U ) and H(V ) finite over
k, then
cxA (U, V ) = cxA (V , U ) .
Proof. The desired assertion results from the following chain of equalities:
cxA (U, V ) = cxA (U, V ∗∗ )
= cxA (U ⊗k V ∗ , k)
= cxA (k, U ⊗k V ∗ )
= cxA (k, Homk (V , U ))
= cxA (V , U ) .
The first and fourth ones come, respectively, from isomorphisms V V ∗∗
and U ⊗k V ∗ Homk (V , U ) given by Lemma 4.8. The second and fifth ones
one follow from the adjunction isomorphism (4.1.3). Theorem 4.7 supplies the
middle link.
We have not yet provided any non-trivial example of DG Hopf k-algebra
A with finitely generated cohomology algebra. To state a general result in
this direction, recall that A is said to be cocommutative if its comultiplication
satisfies the equality = τ , where τ : A ⊗k A → A ⊗k A is defined by
τ (a ⊗ b) = (−1)|a||b| b ⊗ a.
The following result is Wilkerson’s main theorem in [17]:
4.10. If A is a cocommutative DG Hopf k-algebra with zero differential

and with rankk A finite, then the graded k-algebra S = ExtA (k, k) is finitely
generated.
In positive characteristic the proof of the theorem depends on the action of

the Steenrod algebra on S; the existence of such an action is another non-trivial
result, see [11]. The situation is completely different in characteristic zero,
where a celebrated theorem of Hopf shows that every non-negative cocommu-
tative graded Hopf algebra is isomorphic, as an algebra, to the exterior algebra
on a vector space of finite rank; in this case the cohomology is well known, and
is computed next.
5. Exterior algebras I.
In this section k is a field and a DG algebra with ∂ = 0 and with
an exterior k-algebra on alternating indeterminates ξ1 , . . . , ξc of positive odd
degrees.
Proposition 5.1. The graded k-algebra S = Ext∗ (k, k) is a polynomial ring

over k on commuting indeterminates χ1 , . . . , χc with |χi | = −(|ξi | + 1).
Proof. One has ∼ = kξ1 ⊗k B, where kξj denotes the exterior algebra on
the graded vector space kξi and B = kξ2 ⊗k · · · ⊗k kξc . By Lemma 3.3
and induction, it suffices to treat the case = kξ . We use the notation from
Construction 3.1.
Set d = |ξ |, and let F be a DG -module, whose underlying graded -
module has a basis {en }n0 with |en | = (d + 1)n, and with differential defined
by the formulas ∂(e0 ) = 0 and ∂(en ) = ξ en−1 for n ≥ 1. An elementary verifi-
cation shows that the -linear map ε : F → k with ε(e0 ) = 1 and ε(en ) = 0 for
n ≥ 1 is a quasi-isomorphism of DG -modules, and thus a semifree resolution
of k. It yields
Extn (k, k) = Hn (Hom (F, k)) = Hom (F, k)n

kχ (i) for n = (d + 1)i ≥ 0 ;
=
0 otherwise ,
where χ (i) : F → k is the chain map of DG -modules defined by χ (i) (en ) = 0

for n = i and χ (i) (ei ) = 1. Setting χ
(i) (en ) = en−i for n ≥ i and χ (i) (en ) = 0
for n < i one obtains a chain map χ : F → F , such that ε
(i)
χ = χ (i) . The
(i)
definition of composition products yields equalities
[χ (j ) ][χ (i) ] = [χ (j ) ◦ χ
(i) ] = [χ (i+j ) ] .
for all i, j ≥ 0. They show that the isomorphism of graded k-vector spaces
k[χ ] → S, which sends χ i to [χ (i) ] for each i ≥ 0, is a homomorphism of
graded k-algebras.
Remark 5.2. The universal property of exterior algebras guarantees that there
is a unique homomorphism of graded k-algebras : → ⊗k with
(ξi ) = ξi ⊗ 1 + 1 ⊗ ξi for 1 ≤ i ≤ c.
It evidently satisfies (εA ⊗ A) = idA = (A ⊗k εA ), and so is the comulti-

plication of a graded Hopf algebra structure on .
In view of the preceding proposition and remark, Theorems 4.7 and 4.9
yield:
Theorem 5.3. For DG -modules U, V with H(U ) and H(V ) finite over k one
has
cxA (V , U ) = cxA (U, V ) ≤ cx (U, U ) = inj cx U = cx U ≤ cx k = c .
6. Complete intersection local rings I.

In this section (R, m, k) denotes a local ring. When R is complete intersection
(the definition is recalled below) we open a path to an exterior algebra, and use
it to transport results from Section 5.
The embedding dimension of (R, m, k) is the number edim R =
rankk (m/m2 ); by Nakayama’s Lemma, it is equal to the minimal number of
generators of m.
Construction 6.1. Choose a minimal set of generators r1 , . . . , re of m, and let

K denote the Koszul complex on r1 , . . . , re . The functor − ⊗R K preserves
quasi-isomorphisms of complexes of R-modules, so it defines an exact functor
k : D(R) → D(K) .
Lemma 6.2. For each homologically finite complex of R-modules M one has:
cxR M = cxK k(M) and inj cxR M = inj cxK k(M) .

Proof. Let G → M be a semifree resolution over R, and note that the

induced morphism G ⊗R K → M ⊗R K then is a semifree resolution over
K. The expression for cxR M now follows from the isomorphisms of k-vector
spaces
ExtnK (M ⊗R K, k) = Hn (HomK (G ⊗R K, k))

∼
= Hn (HomR (G, k))
= ExtnR (M, k) .
For the second equality, choose a semifree resolution F → k over K. One

has an isomorphism kM HomR (K, Σe M) of DG K-modules, where e =
edim R. For each n ∈ Z adjunction gives the isomorphism of k-vector spaces
below, and the last equality holds because F is semifree over R:
ExtnK (k, M ⊗R K) = Hn (HomK (F, HomR (K, Σe M)))

∼ Hn (HomR (F, Σe M))
=
= Extn−e
R (k, M) .
Krull’s Principal Ideal Theorem implies edim R ≥ dim R, where dim R

denotes the Krull dimension of R. The codimension of R is the number
codim R = edim R − dim R. Rings of codimension zero are called regular.
Cohen’s Structure Theorem shows that the m-adic completion R of R admits

a surjective homomorphism of rings κ : (Q, q, k) → R, with Q a regular local
ring; see [13, §29]. Such a homomorphism is called a Cohen presentation of
One has dim Q ≥ edim R, and equality is equivalent to Ker(κ) ⊆ q2 ; such
R.
a presentation is said to be minimal.
Any Cohen presentation κ can be refined to a minimal one. Indeed, choose
elements q1 , . . . , qn in q mapping to a k-basis of (Ker(κ) + q2 )/q2 . One then
has
codim Q/(q) = (edim Q − n) − (dim Q − n) = codim Q = 0 ,
so the local ring Q/(q) is regular along with Q. The map κ factors through a
homomorphism Q/(q) → R, and the latter is a minimal Cohen presentation.
Construction 6.3. Let ι : R → R be the completion map and choose a minimal

Cohen presentation κ : (Q, q, k) → R.
Choose a minimal set of generators r1 , . . . , re of m and then pick in Q ele-
ments q1 , . . . , qe with κ(qi ) = ι(ri ) for i = 1, . . . , e. Let K denote the Koszul
complex on r1 , . . . , re , and E be the one on q1 , . . . , qe . The definition of Koszul
complexes allows one to identify the DG algebras R ⊗R K and R ⊗Q E.
Choose a minimal set of generators {f1 , . . . , fc } of the ideal Ker κ, let A

be the Koszul complex on this set, and let π : A → R denote the canonical
projection.
Set = A ⊗Q k, and note that this is a DG algebra with zero differential,

and its underlying graded algebra is the exterior algebra k (A1 ⊗Q k).
The ring R is said to be complete intersection if in some Cohen presentation
κ: Q → R the ideal Ker κ can be generated by a Q-regular sequence. When
this is the case, the kernel of every Cohen presentation is generated by a
regular sequence, and for each minimal presentation such a sequence consists
of codim R elements.
In the next three lemmas (R, m, k) denotes a complete intersection local
ring.
Lemma 6.4. The following maps are quasi-isomorphisms of DG algebras:
π⊗ E A⊗Q εE
K = R ⊗R K
ι⊗R K
/R ⊗Q E o Q A ⊗Q E
⊗R K = R / A ⊗Q k = .
is flat over R, one can identify H(ι ⊗R K) with the map

Proof. As R
⊗R H(K) ,
ι ⊗R H(K) : R ⊗R H(K) → R
which is bijective because H(K) is a direct sum of shifts of k.
The sequence f1 , . . . , fc is Q-regular because R is complete intersection,
hence π is a quasi-isomorphism, and then so is π ⊗Q E.
Since the ring Q is regular and the elements q1 , . . . , qe minimally generate
q, they form a Q-regular sequence, so εE is a quasi-isomorphism, and then so
is A ⊗Q εE .
In view of 2.4, the quasi-isomorphisms in Lemma 6.4 define an equivalence
of categories j : D(K) → D(). Lemmas 6.2 and 2.5 then give:
Lemma 6.5. For every homologically finite complex of R-modules M one has
cxR M = cx jk(M) and inj cxR M = inj cx jk(M) .
c
Lemma 6.6. In D() there is an isomorphism jk(k) i 0 Σi k (i ) .
Proof. Choose bases x1 , . . . , xc of A1 over Q with ∂(xi ) = fi for 1 ≤ i ≤ c and

y1 , . . . , ye of E1 over Q with ∂(yj ) = qj for 1 ≤ j ≤ e; Thus fi = dj=1 bij qj
with bij ∈ Q. There is a unique homomorphism α : A → E of DG algebras
over Q with

e
α(xi ) = bij yj for 1 ≤ i ≤ c .
j =1
It appears in a commutative diagram of DG algebras

π⊗ E A⊗Q εE
K R⊗R K
ι⊗R K
/R ⊗Q E o Q A ⊗ Q E
⊗R K R / A ⊗Q k

ε R ⊗R K εR ⊗Q E ε R ⊗Q E α⊗Q E α⊗Q k
ε E
⊗Q E E⊗Q ε E
k(k) k⊗R K k⊗R K k⊗Q E o
E ⊗Q E
/ E ⊗Q k
where εE ⊗Q E and E ⊗Q εE are quasi-isomorphisms because εE is one.

The map
E ⊗R K defines the action of K on k(k), and α ⊗Q k that of
on E ⊗Q k. The commutativity of the diagram implies and jk(k) E ⊗Q k in
D(), see 2.4. The minimality of the Cohen presentation κ means that each bij

is in q, so one has k(k) ∼
c
= i 0 Σi k (i ) as DG K-modules, and hence in D()
one gets

i (ci) c c
jk(k) j Σk Σi j(k)(i ) Σi k ( i ) ,
i 0 i 0 i 0
where the last isomorphism holds because one has j(k) k by Lemma 2.5.
We come to the main result of this section. All its assertions are known: the
first equality is proved in [2, 5.3], the inequality follows from [8, 4.2], and the
second equality from [16, Thm. 6]. The point here is that they are deduced, in
a uniform way, from the corresponding relations for DG modules over exterior
algebras, established in Theorem 5.3, which ultimately are much simpler to
prove.
Theorem 6.7. If (R, m, k) is complete intersection, then for every complex of
R-modules M with H(M) finitely generated the following inequalities hold:
inj cxR M = cxR M ≤ cxR k = codim R .
Proof. The isomorphism of Lemma 6.6 implies cx k = cx jk(k). Now
Lemma 6.5 translates part of Theorem 5.3 into the desired statement.
The equalities in the theorem may fail when R is not complete intersection:
Remark 6.8. Gulliksen [9, 2.3] proved that the condition cxR k < ∞ charac-
terizes local complete intersection rings among all local rings.
Jorgensen and Şega [10, 1.2] construct Gorenstein local k-algebras R with
rankk R finite and modules M with {cxR M, inj cxR M} = {1, ∞}, in one order
or the other.
Remark 6.9. The unused portion of Theorem 5.3 suggests that the relations
cxR (N, M) = cxR (M, N) ≤ cxR (M, M) = cxR M
might hold also over complete intersections. They do, see [4, 5.7], but we know
of no simple way to deduce them from Theorem 5.3.
7. Projective levels
In this section A is a DG algebra. We collect some results on A-levels of
DG modules, which are reminiscent of theorems on projective dimension for
modules.
7.1. For any DG A-module U the following hold.
(1) One has levelA A (U ) ≤ l + 1 if and only if U is isomorphic in D(A) to a
direct summand of some semifree DG module F with a semifree filtration
having F l = F and every F n /F n−1 of finite rank over A ; see 2.1.
(2) When A is non-negative, H0 (A) is a field, and Hi (U ) = 0 for all i 0,
one has levelAA (U ) < ∞ if and only if Tor (k, U ) is finite over k.
A
(3) When A has zero differential and is noetherian, and the graded A-module
H(U ) is finitely generated, the following inequalities hold:
A (U ) ≤ proj dimA H(U ) + 1 ≤ gl dim A + 1 .

levelA
Indeed, (1) is proved in [5, 4.2], (2) in [5, 4.8], and (3) in [5, 5.5].
The preceding results pertain to homological algebra, in the sense that they

do not depend on the structure of the ring A = n∈Z An . For algebras over
fields the next result contains as a special case the New Intersection Theorem,
see [14], a central result in commutative algebra, and effectively belongs to the
latter subject.
7.2. Let A be a DG algebra with zero differential and U a DG A-module.
If the ring A is commutative, noetherian, and is an algebra over a field, then
levelAA (U ) ≥ height I + 1 ,

where I is the annihilator of the A -module n∈Z Hn (U ); see [5, 5.1].
8. Exterior algebras II.

In this section k is a field, and a DG algebra with ∂ = 0 and an exterior
k-algebra on alternating indeterminates of positive odd degrees d1 , . . . , dc .
We recall a version of the Bernstein-Gelfand-Gelfand equivalence from [5,
7.4]:
8.1. Let S be a DG algebra with ∂ S = 0 and S a polynomial ring over k

on commuting indeterminates of degrees −(d1 + 1), . . . , −(dc + 1); set d =
d1 + · · · + dc .
h t
There exist exact functors D() −
→ D(S) −
→ D() inducing inverse equiva-
lences
h /
thick (k) o ≡ thickS (S)
t
| | (8.1.1)
h /
thick () o ≡ thickS (k)
t
such that in D(S) and D(), respectively, there are isomorphisms

h() Σd k t(k) Σ−d
and (8.1.2)
h(k) S t(S) k .
Let Df () be the full subcategory of D() whose objects are the DG modules
U with H(U ) finite over , and Df (S) the corresponding subcategory of D(S).
The next proposition refines the equivalences in (8.1.1); see Remark 8.3.
Proposition 8.2. For each U ∈ Df () and each M ∈ Df (S) one has
cx U = dimS H(hU ) and dimS M = cx H(tM) ,
and for each n ∈ N the functors h and t induce inverse equivalences

h /
Df () o ≡ Df (S)
t
| | (8.2.1)
h /
{U ∈ Df () | cx U ≤ n} o ≡ {M ∈ Df (S) | dimS H(M) ≤ n} .
t
Proof. One has Df () = thick (k) by 1.4(1). As the polynomial ring S has
finite global dimension, 7.1(3) gives Df (S) = thickS (S). Thus, the top row of
(8.1.1) gives inverse equivalences between Df () and Df (S). It is easy to verify
that the subcategories in the lower row are thick, so it suffices to compute cx U
and dimS H(M).
The equivalence h gives the first isomorphism in the chain
RHom (k, U ) RHomS (h(k), h(U )) RHomS (S, h(U )) h(U ) ;
the second one comes from (8.1.2), the third is clear. In homology, one obtains
Extn (k, U ) ∼
= Hn (h(U ))
for every n ∈ Z, whence the second equality in the following sequence; the
first one comes from Theorem 5.3, the third from the Hilbert-Serre Theorem,
see 4.6:
cx U = inj cx U = cx ( rankk Hn (h(U )))n∈Z = dimS H(h(U )) .
Finally, from the already proved assertions one obtains
dimS H(M) = dimS H(ht(M)) = cx t(M) .
Remark 8.3. For U ∈ Df () the definition of complexity shows that cx U ≤ 0
is equivalent to the finiteness of the k-vector space Ext (U, k). It is the graded
k-dual of Tor (U, k), so by 7.1(2) its finiteness is equivalent to U ∈ thick ().
On the other hand, M ∈ Df (S) has dimS H(M) ≤ 0 if and only if M is in
thickS (k); see 1.4(1). Thus, for n = 0 diagram (8.2.1) reduces to (8.1.1).
Theorem 8.4. If U is a DG -module with 0 < rankk H(U ) < ∞, then one
has
card{n ∈ Z | Hn (U ) = 0} ≥ levelk (U ) ≥ c − cx U + 1 .
Proof. The inequality on the left comes from 1.4(3). The one on the right results
from the following chain of (in)equalities
levelk (U ) = levelhS(k) (h(U ))

= levelSS (h(U ))
≥ height(AnnS H(h(U ))) + 1
= dim S − dimS H(h(U )) + 1
= dim S − cx U + 1
= c − cx V + 1 .
The first one holds because h is an equivalence, see 1.2(2), the second comes
from (8.1.2), and the third from 7.2; the remaining equalities hold by [13,
Exercise 5.1]1 , by Lemma 8.2, and because dim S = c holds.
The inequalities in Theorem 8.4 are tight.
Example 8.5. For each integer i with 0 ≤ i ≤ c and (i) = /(ξ1 , . . . , ξi ) one
has
card{n ∈ Z | Hn ((i) ) = 0} = c − i + 1 and cxR (i) = i .
1 solved on page 288

Indeed, the first one is clear. For (i) = /(ξi+1 , . . . , ξn ) one has
Extn ((i) , k) ∼
= Extn(i) ⊗k (i) ((i) , k) ∼
= Extn(i) (k, k) ,
because is isomorphic to (i) ⊗k (i) as k-algebras. Theorem 5.3 gives

cx(i) k = i.
9. Complete intersection local rings II.

We finish the paper with a new proof of [5, 11.3]. The original argument
uses a reduction, constructed in [2], to a complete intersection ring of smaller
codimension and a bounded complex of free modules over it, then refers to the
main theorem of [5], which applies to such complexes over arbitrary local rings.
Here we describe a direct reduction to results on exterior algebras, established
in Section 8.
Theorem 9.1. If (R, m, k) is a complete intersection local ring and M a

complex of R-modules with H(M) finite and nonzero, then one has inequalities

R Hn (M) ≥ levelkR (M) ≥ codim R − cxR M + 1 .
n∈Z
Proof. The inequality on the left is a special case of 1.4(2).

Set c = codim R. Let K be the Koszul complex on a minimal set of gen-
erators for m and the DG algebra with zero differential, and with under-
lying graded algebra an exterior algebra over k on c generators of degree
one. Construction 6.1 and Lemma 6.4 provide exact functors of triangulated
categories
D(R)
k / D(K) j
/ D()
≡
the second of which is an equivalence. One then has the following relations
jk(k)
levelkR (M) ≥ level (jk(M))
= levelk (jk(M))
≥ c − cx (jk(M)) + 1
= codim R − cxR M + 1
where the inequalities are given by 1.2(2) and by Theorem 8.4, while the
equalities come from 1.2(1) and Lemmas 6.6, and from Lemma 6.5.
As in Theorem 8.4 the inequalities in Theorem 9.1 are tight.

Example 9.2. Let k be a field and set R = k[x1 , . . . , xc ]/(x12 , . . . , xc2 ). For
every integer i with 0 ≤ i ≤ c, set R (i) = R/(x1 , . . . , xi ). One then has equal-
ities
R Ri = c − i + 1 , codim R = c , and cxR R (i) = i .
Indeed, the first two are clear. For Q(i) = R/(xi+1 , . . . , xn ) one has
ExtnR (R (i) , k) ∼
= ExtnQ(i) ⊗k R(i) (R (i) , k) ∼
= ExtnQ(i) (k, k) ,
because R is isomorphic to Q(i) ⊗k R (i) as k-algebras. Theorem 6.7 gives
cxQ(i) k = i.
The number on the right hand side of the formula in Theorem 9.1 is finite
whenever the module M has finite complexity. However, this condition alone
does not imply the inequalities in the theorem, even when the ring R is
Gorenstein.
Example 9.3. Let k be a field and c an integer with c ≥ 3. The ring
k[x1 , . . . , xc ]
R = c−1
2
h=1 (xh − xh+1
2
) + 1i<j c (xi xj )
is Gorenstein, but not complete intersection, and its module M = R has
R M = levelkR (M) = 3 < c − 0 + 1 = codim R − cxR M + 1 .
It is worth noting that a very special case of Theorem 9.1 was initially
discovered when studying actions of finite elementary abelian groups on finite
CW complexes:
Remark 9.4. Let k be a field of positive characteristic p and G an ele-
mentary abelian p-group of rank c. The group algebra kG is isomorphic to
p p
k[x1 , . . . , xc ]/(x1 , . . . , xc ), which is a complete intersection of codimension
c. Over kG Theorem 9.1 was proved by Carlsson [7] for M of finite rank over
A and p equal to 2, and for general M and p by Allday, Baumgartner, and
Puppe; see [1, 4.6.42].
References
[1] C. Allday, V. Puppe, Cohomological methods in transf ormation groups, Cam-
bridge Stud. Adv. Math. 32, Cambridge Univ. Press, Cambridge, 1993.
[2] L. L. Avramov, Modules of finite virtual projective dimension, Invent. Math. 96
(1989), 71–101.
[3] L. L. Avramov, Homological asymptotics of modules over local rings, Commuta-

tive algebra (Berkeley, CA, 1987), Math. Sci. Res. Inst. Publ., 15, Springer, New
York, 1989; 33–62.
[4] L. L. Avramov, R.-O. Buchweitz, Support varieties and cohomology over complete
intersections, Invent. Math. 142 (2000), 285–318.
[5] L. L. Avramov, R.-O. Buchweitz, S. B. Iyengar, C. Miller, Homology of finite free
complexes, Adv. Math. 223 (2010), 1731–1781.
[6] L. L. Avramov, S. Halperin, Through the looking glass: a dictionary between
rational homotopy theory and local algebra, Algebra, algebraic topology and their
interactions (Stockholm, 1983), Lecture Notes in Math. 1183, Springer, Berlin,
1986; 1–27
[7] G. Carlsson, On the homology of finite free (Z/2)k -complexes, Invent. Math. 74
(1983), 139–147.
[8] T. H. Gulliksen, A change of ring theorem with applications to Poincaré series
and intersection multiplicity, Math. Scand. 34 (1974), 167–183.
[9] T. H. Gulliksen, On the deviations of a local ring, Math. Scand. 47 (1980), 5–20.
[10] D. A. Jorgensen, L. M. Şega, Asymmetric complete resolutions and vanishing of
Ext over Gorenstein rings, Int. Math. Res. Not. 2005, 3459–3477.
[11] A. Liulevicius, The factorization of cyclic reduced powers by secondary cohomol-
ogy operations, Mem. Amer. Math. Soc. 42, Amer. Math. Soc., Providence, RI,
1962.
[12] S. Maclane, Homology, Grundlehren math. Wissenschaften, 114, Springer, Berlin,
1963.
[13] H. Matsumura, Commutative ring theory, Cambridge Stud. Adv. Math. 8 Cam-
bridge Univ. Press, Cambridge, 1986.
[14] P. C. Roberts, Multiplicities and Chern classes in local algebra, Cambridge Tracts
Math. 133, Cambridge Univ. Press, Cambridge, 1998.
[15] W. Smoke, Dimension and multiplicity over graded algebras, J. Algebra 21 (1972),
149–173.
[16] J. Tate, Homology of noetherian rings and local rings, Illinois J. Math. 1 (1957),
14–25
[17] C. Wilkerson, The cohomology algebras of finite-dimensional Hopf algebras,
Trans. Amer. Math. Soc. 264 (1981), 137–150.
Department of Mathematics, University of Nebraska, Lincoln, NE 68588, U.S.A.

E-mail address:

Department of Mathematics, University of Nebraska, Lincoln, NE 68588, U.S.A.

E-mail address:

Cluster algebras, quiver representations and
triangulated categories
bernhard keller
Abstract. This is an introduction to some aspects of Fomin-Zelevinsky’s clus-
ter algebras and their links with the representation theory of quivers and with
Calabi-Yau triangulated categories. It is based on lectures given by the author at
summer schools held in 2006 (Bavaria) and 2008 (Jerusalem). In addition to by
now classical material, we present the outline of a proof of the periodicity con-
jecture for pairs of Dynkin diagrams (details will appear elsewhere) and recent
results on the interpretation of mutations as derived equivalences.
Contents
1. Introduction 76
2. An informal introduction to cluster-finite cluster algebras 79
3. Symmetric cluster algebras without coefficients 83
4. Cluster algebras with coefficients 92
5. Categorification via cluster categories: the finite case 99
6. Categorification via cluster categories: the acyclic case 122
7. Categorification via 2-Calabi-Yau categories 126
8. Application: The periodicity conjecture 138
9. Quiver mutation and derived equivalence 143
References 154
1. Introduction
1.1. Context
Cluster algebras were invented by S. Fomin and A. Zelevinsky [50] in the
spring of the year 2000 in a project whose aim it was to develop a combinato-
rial approach to the results obtained by G. Lusztig concerning total positivity in
algebraic groups [103] on the one hand and canonical bases in quantum groups
[102] on the other hand (let us stress that canonical bases were discovered
Date: July 2008, last modified on May 13, 2010.
76
Cluster algebras, representations, triangulated categories 77
independently and simultaneously by M. Kashiwara [83]). Despite great

progress during the last few years [52] [17] [55], we are still relatively far
from these initial aims. Presently, the best results on the link between cluster
algebras and canonical bases are probably those of C. Geiss, B. Leclerc and
J. Schröer [64] [65] [62] [61] [63] but even they cannot construct canonical
bases from cluster variables for the moment. Despite these difficulties, the
theory of cluster algebras has witnessed spectacular growth thanks notably
to the many links that have been discovered with a wide range of subjects
including
r Poisson geometry [69] [70] . . . ,
r integrable systems [54] . . . ,
r higher Teichmüller spaces [43] [44] [45] [46] . . . ,
r combinatorics and the study of combinatorial polyhedra like the Stasheff
associahedra [34] [33] [99] [48] [108] [49] . . . ,
r commutative and non commutative algebraic geometry, in particular the
study of stability conditions in the sense of Bridgeland [23] [21] [24],
Calabi-Yau algebras [71] [35], Donaldson-Thomas invariants [123] [95] [96]
[98] . . . ,
r and last not least the representation theory of quivers and finite-dimensional
algebras, cf. for example the surveys [9] [114] [116] .
We refer to the introductory papers [53] [132] [134] [135] [136] and to the
cluster algebras portal [47] for more information on cluster algebras and their
links with other parts of mathematics.
The link between cluster algebras and quiver representations follows the
spirit of categorification: One tries to interpret cluster algebras as combinatorial
(perhaps K-theoretic) invariants associated with categories of representations.
Thanks to the rich structure of these categories, one can then hope to prove
results on cluster algebras which seem beyond the scope of the purely combina-
torial methods. It turns out that the link becomes especially beautiful if we use
triangulated categories constructed from categories of quiver representations.
1.2. Contents
We start with an informal presentation of Fomin-Zelevinsky’s classification the-
orem and of the cluster algebras (without coefficients) associated with Dynkin
diagrams. Then we successively introduce quiver mutations, the cluster algebra
associated with a quiver, and the cluster algebra with coefficients associated
with an ‘ice quiver’ (a quiver some of whose vertices are frozen). We illustrate
78 Bernhard Keller
cluster algebras with coefficients on a number of examples appearing as coor-

dinate algebras of homogeneous varieties.
Sections 5, 6 and 7 are devoted to the (additive) categorification of cluster
algebras. We start by recalling basic notions from the representation theory of
quivers. Then we present a fundamental link between indecomposable repre-
sentations and cluster variables: the Caldero-Chapoton formula. After a brief
reminder on derived categories in general, we give the canonical presenta-
tion in terms of generators and relations of the derived category of a Dynkin
quiver. This yields in particular a presentation for the module category, which
we use to sketch Caldero-Chapoton’s proof of their formula. Then we intro-
duce the cluster category and survey its many links to the cluster algebra in
the finite case. Most of these links are still valid, mutatis mutandis, in the
acyclic case, as we see in section 6. Surprisingly enough, one can go even
further and categorify interesting classes of cluster algebras using general-
izations of the cluster category, which are still triangulated categories and
Calabi-Yau of dimension 2. We present this relatively recent theory in sec-
tion 7. In section 8, we apply it to sketch a proof of the periodicity conjecture
for pairs of Dynkin diagrams (details will appear elsewhere [86]). In the final
section 9, we give an interpretation of quiver mutation in terms of derived
equivalences. We use this framework to establish links between various ways
of lifting the mutation operation from combinatorics to linear or homological
algebra: mutation of cluster-tilting objects, spherical collections and decorated
representations.
Acknowledgments
These notes are based on lectures given at the IRTG-Summerschool 2006
(Schloss Reisensburg, Bavaria) and at the Midrasha Mathematicae 2008
(Hebrew University, Jerusalem). I thank the organizers of these events for
their generous invitations and for providing stimulating working conditions. I
am grateful to Thorsten Holm, Peter Jørgensen and Raphael Rouquier for their
encouragment and for accepting to include these notes in the proceedings of
the ‘Workshop on triangulated categories’ they organized at Leeds in 2006.
It is a pleasure to thank to Carles Casacuberta, André Joyal, Joachim Kock,
Amnon Neeman and Frank Neumann for an invitation to the Centre de Recerca
Matemàtica, Barcelona, where most of this text was written down. I thank
Lingyan Guo and Sefi Ladkani for kindly pointing out misprints and inaccu-
racies. I am indebted to Tom Bridgeland, Bernard Leclerc, David Kazhdan,
Tomoki Nakanishi, Raphaël Rouquier and Michel Van den Bergh for helpful
conversations.
2. An informal introduction to cluster-finite cluster algebras

2.1. The classification theorem
Let us start with a remark on terminology: a cluster is a group of similar things
or people positioned or occurring closely together [121], as in the combination
‘star cluster’. In French, ‘star cluster’ is translated as ‘amas d’étoiles’, whence
the term ‘algèbre amassée’ for cluster algebra.
We postpone the precise definition of a cluster algebra to section 3. For the
moment, the following description will suffice: A cluster algebra is a com-
mutative Q-algebra endowed with a family of distinguished generators (the
cluster variables) grouped into overlapping subsets (the clusters) of fixed finite
cardinality, which are constructed recursively using mutations.
The set of cluster variables in a cluster algebra may be finite or infinite. The
first important result of the theory is the classification of those cluster algebras
where it is finite: the cluster-finite cluster algebras. This is the
Classification Theorem 2.1 (Fomin-Zelevinsky [52]). The cluster-finite clus-

ter algebras are parametrized by the finite root systems (like semisimple com-
plex Lie algebras).
It follows that for each Dynkin diagram , there is a canonical cluster

algebra A . It turns out that A occurs naturally as a subalgebra of the field of
rational functions Q(x1 , . . . , xn ), where n is the number of vertices of . Since
A is generated by its cluster variables (like any cluster algebra), it suffices
to produce the (finite) list of these variables in order to describe A . Now for
the algebras A , the recursive construction via mutations mentioned above
simplifies considerably. In fact, it turns out that one can directly construct the
cluster variables without first constructing the clusters. This is made possible by
2.2. The knitting algorithm

The general algorithm will become clear from the following three examples.
We start with the simplest non trivial Dynkin diagram:
= A2 : ◦ ◦.
We first choose a numbering of its vertices and an orientation of its edges:

= A2 : 1
/2.
associated
Now we draw the so-called repetition (or Bratteli diagram) Z
made up of a countable number of
with : We first draw the product Z ×
80 Bernhard Keller
copies of (drawn slanted upwards); then for each arrow α : i → j of , we

add a new family of arrows (n, α ∗ ) : (n, j ) → (n + 1, i), n ∈ Z (drawn slanted
downwards). We refer to section 5.5 for the formal definition. Here is the result
= A2 :
for
... ? ◦ >> ? ◦ ??? ? ◦ ??? ? ◦ >>> ? ◦ ...
>
◦ ◦ ◦ ◦ ◦
We will now assign a cluster variable to each vertex of the repetition. We
start by assigning x1 and x2 to the vertices of the zeroth copy of . Next, we

construct new variables x1 , x2 , x1 , . . . by ‘knitting’ from the left to the right
(an analogous procedure can be used to go from the right to the left).
... ? x2 > x2 x2 ? ◦ >> ?◦ ...
? ?? ? ??? >>
>> ?

x1 x1 x1 x1 ◦
To compute x1 , we consider its immediate predecessor x2 , add 1 to it and divide

the result by the left translate of x1 , to wit the variable x1 . This yields
1 + x2
x1 = .
x1
Similarly, we compute x2 by adding 1 to its predecessor x1 and dividing the
result by the left translate x2 :
1 + x1 x1 + 1 + x2
x2 = = .
x2 x1 x2
Using the same rule for x1 we obtain

1 + x2 x1 x2 + x1 + 1 + x2 1 + x2 1 + x1
x1 = = / = .
x1 x1 x2 x1 x2
Here something remarkable has happened: The numerator x1 x2 + x1 + 1 + x2
is actually divisible by 1 + x2 so that the denominator remains a monomial
(contrary to what one might expect). We continue with

1 + x1 x2 + 1 + x1 x1 + 1 + x2
x2 = = / = x1 ,
x2 x2 x1 x2
a result which is perhaps even more surprising. Finally, we get

1 + x2 1 + x1
x1 = = (1 + x1 )/ = x2 .
x1 x2
Clearly, from here on, the pattern will repeat. We could have computed ‘towards
the left’ and would have found the same repeating pattern. In conclusion, there
are the 5 cluster variables x1 , x2 , x1 , x2 and x1 and the cluster algebra AA2 is the
Q-subalgebra (not the subfield!) of Q(x1 , x2 ) generated by these 5 variables.
Before going on to a more complicated example, let us record the remarkable
phenomena we have observed:
(1) All denominators of all cluster variables are monomials. In other words,
the cluster variables are Laurent polynomials. This Laurent phenomenon
holds for all cluster variables in all cluster algebras, as shown by Fomin
and Zelevinsky [51].
(2) The computation is periodic and thus only yields finitely many cluster
variables. Of course, this was to be expected by the classification theorem
above. In fact, the procedure generalizes easily from Dynkin diagrams to
arbitrary trees, and then periodicity characterizes Dynkin diagrams among
trees.
(3) Numerology: We have obtained 5 cluster variables. Now we have 5 = 2 + 3
and this decomposition does correspond to a natural partition of the set of
cluster variables into the two initial cluster variables x1 and x2 and the
three non initial ones x1 , x2 and x1 . The latter are in natural bijection with
the positive roots α1 , α1 + α2 and α2 of the root system of type A2 with
simple roots α1 and α2 . To see this, it suffices to look at the denominators
of the three variables: The denominator x1d1 x2d2 corresponds to the root
d1 α1 + d2 α2 . It was proved by Fomin-Zelevinsky [52] that this generalizes
to arbitrary Dynkin diagrams. In particular, the number of cluster variables
in the cluster algebra A always equals the sum of the rank and the number
of positive roots of .
Let us now consider the example A3 : We choose the following linear orientation:
1 /2 / 3.
The associated repetition looks as follows:

◦? ? x3 >> x3 x1
x
?? > ? >>
> ? >>> ? 3 ??
?
? >
x2 x2 x2 x x
~? ??? ? @@@ ? ???? ~~? 2
2
~ ? @@@
~~
~
x1 x1 x1 x1 x3
The computation of x1 is as before:

1 + x2
x1 = .
x1
However, to compute x2 , we have to modify the rule, since x2 has two immediate
predecessors with associated variables x1 and x3 . In the formula, we simply
82 Bernhard Keller
take the product over all immediate predecessors:

1 + x1 x3 x1 + x3 + x2 x3
x2 = = .
x2 x2 x3
Similarly, for the following variables x3 , x1 , . . . . We obtain the periodic pattern
shown in the diagram above. In total, we find 9 = 3 + 6 cluster variables,
namely
1 + x2 x1 + x3 + x2 x3 x1 + x1 x2 + x3 + x2 x3
x1 , x2 , x3 , , , ,
x1 x1 x2 x1 x2 x3
x1 + x3 x1 + x1 x2 + x3 1 + x2
, , .
x2 x2 x3 x3
The cluster algebra AA3 is the subalgebra of the field Q(x1 , x2 , x3 ) generated by
these variables. Again we observe that all denominators are monomials. Notice
also that 9 = 3 + 6 and that 3 is the rank of the root system associated with A3
and 6 its number of positive roots. Moreover, if we look at the denominators
of the non initial cluster variables (those other than x1 , x2 , x3 ), we see a natural
bijection with the positive roots
α1 , α1 + α2 , α1 + α2 + α3 , α2 , α2 + α3 , α3
of the root system of A3 , where α1 , α2 , α3 denote the three simple roots.
Finally, let us consider the non simply laced Dynkin diagram = G2 :
(3,1)
◦ ◦.
The associated Cartan matrix is

2 −3
−1 2
and the associated root system of rank 2 looks as follows:
3α1 +O 2α2
α2 gOO α +α 2α1D + α2 3α 1 + α2
OOO 1 Z44 2
o o7
OOO 44
o
OOO 4
ooo
OOO44
ooooo
O
oo
−α1 o o o
4O4OOO / α1
o o o
4 O O
ooo
44 OOO
ooo 44 OOO
oo
4 OO'
o o
wo
−α2

We choose an orientation of the valued edge of G2 to obtain the following

valued oriented graph:
: 1

(3,1)
/ 2.
Now the repetition also becomes a valued oriented graph
@ x2555 x2 x2 x2 @ x2

A 77 @ 77 A 55

3,1 1,3
55 3,1 1,3
77 3,1 1,3
77 3,1 1,3
555 3,1

x1 x1 x1 x1 x1
The mutation rule is a variation on the one we are already familiar with: In
the recursion formula, each predecessor p of a cluster variable x has to be
raised to the power indicated by the valuation ‘closest’ to p. Thus, we have for
example
1 + x2 1 + (x1 )3 1 + x13 + 3x2 + 3x22 + x23

x1 = , x2 = = ,
x1 x2 x13 x2
1 + x2 ...
x1 = = 2 ,
x1 x1 x2
where we can read off the denominators from the decompositions of the positive
roots as linear combinations of simple roots given above. We find 8 = 2 + 6
cluster variables, which together generate the cluster algebra AG2 as a subal-
gebra of Q(x1 , x2 ).
3. Symmetric cluster algebras without coefficients

In this section, we will construct the cluster algebra associated with an anti-
symmetric matrix with integer coefficients. Instead of using matrices, we will
use quivers (without loops or 2-cycles), since they are easy to visualize and
well-suited to our later purposes.
3.1. Quivers
Let us recall that a quiver Q is an oriented graph. Thus, it is a quadruple given
by a set Q0 (the set of vertices), a set Q1 (the set of arrows) and two maps
s : Q1 → Q0 and t : Q1 → Q0 which take an arrow to its source respectively
its target. Our quivers are ‘abstract graphs’ but in practice we draw them as in
84 Bernhard Keller
this example:
$ //
Q:
^=== μ
3 α 5 /6
λ ==
==
β
/
1 /2o 4.
ν
γ
A loop in a quiver Q is an arrow α whose source coincides with its target;

a 2-cycle is a pair of distinct arrows β = γ such that the source of β equals
the target of γ and vice versa. It is clear how to define 3-cycles, connected
components. . . . A quiver is finite if both, its set of vertices and its set of arrows,
are finite.
3.2. Seeds and mutations

Fix an integer n ≥ 1. A seed is a pair (R, u), where
a) R is a finite quiver without loops or 2-cycles with vertex set {1, . . . , n};
b) u is a free generating set {u1 , . . . , un } of the field Q(x1 , . . . , xn ) of fractions
of the polynomial ring Q[x1 , . . . , xn ] in n indeterminates.
Notice that in the quiver R of a seed, all arrows between any two given vertices
point in the same direction (since R does not have 2-cycles). Let (R, u) be a
seed and k a vertex of R. The mutation μk (R, u) of (R, u) at k is the seed
(R , u ), where
a) R is obtained from R as follows:
1) reverse all arrows incident with k;
2) for all vertices i = j distinct from k, modify the number of arrows
between i and j as follows:
R R
i:
r /j i ]:
r+st
/j
:: A ::
s t s t
k k
i ]:
r /j i:
r−st
/j
:: :: A
s t s t
k k
where r, s, t are non negative integers, an arrow i

l / j with l ≥ 0 means
that l arrows go from i to j and an arrow i

l / j with l ≤ 0 means that
−l arrows go from j to i.
b) u is obtained from u by replacing the element uk with

⎛ ⎞
1
uk = ⎝ ui + uj ⎠ . (3.2.1)
uk arrows i → k arrows k → j
In the exchange relation (3.2.1), if there are no arrows from i with target k,
the product is taken over the empty set and equals 1. It is not hard to see that
μk (R, u) is indeed a seed and that μk is an involution: we have μk (μk (R, u)) =
(R, u). Notice that the expression given in (3.2.1) for uk is subtraction-free.
To a quiver R without loops or 2-cycles with vertex set {1, . . . , n} there
corresponds the n × n antisymmetric integer matrix B whose entry bij is the
number of arrows i → j minus the number of arrows j → i in R (notice
that at least one of these numbers is zero since R does not have 2-cycles).
Clearly, this correspondence yields a bijection. Under this bijection, the matrix
B corresponding to the mutation μk (R) has the entries

−bij if i = k or j = k;
bij =
bij + sgn(bik )[bik bkj ]+ else,
where [x]+ = max(x, 0). This is matrix mutation as it was defined by Fomin-
Zelevinsky in their seminal paper [50], cf. also [55].
3.3. Examples of seed and quiver mutations

Let R be the cyclic quiver
1 (3.3.1)
E 222
22
22
2
2 o 3
and u = {x1 , x2 , x3 }. If we mutate at k = 1, we obtain the quiver
1
Y333
33
33
3
2 3
and the set of fractions given by u1 = (x2 + x3 )/x1 , u2 = u2 = x2 and u3 =
u3 = x3 . Now, if we mutate again at 1, we obtain the original seed. This is a
general fact: Mutation at k is an involution. If, on the other hand, we mutate
86 Bernhard Keller
(R , u ) at 2, we obtain the quiver
1
E X111
11
11

2 3
and the set u given by u1 = u1 = (x2 + x3 )/x1 , u2 = x1 +x 2 +x3
x1 x2
and u3 = u3 =
x3 .
An important special case of quiver mutation is the mutation at a source (a
vertex without incoming arrows) or a sink (a vertex without outgoing arrows).
In this case, the mutation only reverses the arrows incident with the mutating
vertex. It is easy to see that all orientations of a tree are mutation equivalent and
that only sink and source mutations are needed to pass from one orientation to
any other.
Let us consider the following, more complicated quiver glued together from
four 3-cycles:
1 (3.3.2)
F 22
2
o
E 2 : A 3 33
sg ggg 5 kWWWW3
4 6.
If we successively perform mutations at the vertices 5, 3, 1 and 6, we obtain

the sequence of quivers (we use [87])
1 1 1 1
F 222 E X111 O 222 O 222

2 ]; 3 2 ]; B 3 2 3 2 3

3 WW
o ggg 5 WW+ ggg 3 WWWW+ 3 WW
o ggg 5 WW+ 6 5 kWWWW
4g
g g 5
4 6 4 o 6 4 __________/ 6.
Notice that the last quiver no longer has any oriented cycles and is in fact an
orientation of the Dynkin diagram of type D6 . The sequence of new fractions
appearing in these steps is
x3 x4 +x2 x6 x3 x4 +x1 x5 +x2 x6
u5 = , u3 = ,
x5 x3 x5
x2 x3 x4 +x32 x4 +x1 x2 x5 +x22 x6 +x2 x3 x6 x3 x4 + x4 x5 + x2 x6
u1 = , u6 = .
x1 x3 x5 x5 x6
It is remarkable that all the denominators appearing here are monomials and
that all the coefficients in the numerators are positive.
Finally, let us consider the quiver
E12 (3.3.3)
222

o
F 2 33 F 3 22
33 22

o o
4
F 22 E 5 22 F 6 22
22 22 22

7o 8o 9o 10.
One can show [91] that it is impossible to transform it into a quiver without
oriented cycles by a finite sequence of mutations. However, its mutation class
(the set of all quivers obtained from it by iterated mutations) contains many
quivers with just one oriented cycle, for example
10 \\\- 5
6 } 5 jUUU 3
46 5/
!! /
! ~} Q$$
$
1
7 m\\\ 6 x 6 8
qqq ::
x 7
Z55 |x
7
7$
$$ xx<
2 ooo \99
6
10 NN
8 RRR( lll5 4 X1 10 9 xrr &
> EE" 8* 4#
9! 1 8 UUU* || ** ##
!! 9 1 1
3
3 2.
2
In fact, in this example, the mutation class is finite and it can be completely com-
puted using, for example, [87]: It consists of 5739 quivers up to isomorphism.
The above quivers are members of the mutation class containing relatively few
arrows. The initial quiver is the unique member of its mutation class with the
largest number of arrows. Here are some other quivers in the mutation class
with a relatively large number of arrows:
◦
C -- B ◦- n]]]] ◦ > ◦& jVVVV
- --- E 111 ||| &&& ~? ◦''
◦- m\\\ ◦ m[[[ - 1 ◦& iTTTT ~~ ''
-- C ,,, C ◦. ◦ lYYYYY p````` ◦ && ◦ kWW
◦H 7 G & {{= ''' WW? ◦(
, .. 77 { ' ~ (
◦ mZZZ ◦ n]]]] m[[ 7 ◦ jUUUU ~~ ((
.. C ◦- [ ◦ ◦ qbbbb ◦H 6 ◦ kWWWW
. -- C
H 777 666 ◦ Z4
◦ m\\\ ◦ 7 444
qbbbb ◦ qcccc ◦
◦ ◦ aaaa0 ◦
88 Bernhard Keller
Only 84 among the 5739 quivers in the mutation class contain double
arrows (and none contain arrows of multiplicity ≥ 3). Here is a typical
example
6 1 PPPP
mmmm ' ``````0 4
ww I R%
3I 6

J ww %%%
w2 %
8 X1 {www aaaaaa0
11 10 QaQaQa 2
Q(
7 @5

9
The classification of the quivers with a finite mutation class is still open. Many
examples are given in [49] and [38].
The quivers (3.3.1), (3.3.2) and (3.3.3) are part of a family which appears in
the study of the cluster algebra structure on the coordinate algebra of the sub-
group of upper unitriangular matrices in SL(n, C), cf. section 4.6. The quiver
(3.3.3) is associated with the elliptic root system E8(1,1) in the notations of
Saito [118], cf. Remark 19.4 in [64]. The study of coordinate algebras on vari-
eties associated with reductive algebraic groups (in particular, double Bruhat
cells) has provided a major impetus for the development of cluster algebras,
cf. [17].
3.4. Definition of cluster algebras

Let Q be a finite quiver without loops or 2-cycles with vertex set {1, . . . , n}.
Consider the initial seed (Q, x) consisting of Q and the set x formed by the
variables x1 , . . . , xn . Following [50] we define
r the clusters with respect to Q to be the sets u appearing in seeds (R, u)
obtained from (Q, x) by iterated mutation,
r the cluster variables for Q to be the elements of all clusters,
r the cluster algebra AQ to be the Q-subalgebra of the field Q(x1 , . . . , xn )
generated by all the cluster variables.
Thus, the cluster algebra consists of all Q-linear combinations of monomials

in the cluster variables. It is useful to define yet another combinatorial object
associated with this recursive construction: The exchange graph associated with
Q is the graph whose vertices are the seeds modulo simultaneous renumbering
of the vertices and the associated cluster variables and whose edges correspond
to mutations.
A remarkable theorem due to Gekhtman-Shapiro-Vainshtein states that each
cluster u occurs in a unique seed (R, u), cf. [68].
Notice that the knitting algorithm only produced the cluster variables
whereas this definition yields additional structure: the clusters.
3.5. The example A2

Here the computation of the exchange graph is essentially equivalent to per-
forming the knitting algorithm. If we denote the cluster variables by x1 , x2 , x1 ,
x2 and x1 as in section 2.2, then the exchange graph is the pentagon
(x1 ← x2 )
kk SSS
kkkkkkk SSS
SSS
kk SSS
kkkk SS
(x1 → xG2 ) (x1 → x2 )
GG ww
GG ww
GG w
GG w
ww
(x1 → x1 )
(x1 ← x2 )
where we have written x1 → x2 for the seed (1 → 2, {x1 , x2 }). Notice that it is
not possible to find a consistent labeling of the edges by 1’s and 2’s. The reason
for this is that the vertices of the exchange graph are not the seeds but the seeds
up to renumbering of vertices and variables. Here the clusters are precisely the
pairs of consecutive variables in the cyclic ordering of x1 , . . . , x1 .
3.6. The example A3

Let us consider the quiver
Q: 1 /2 /3
obtained by endowing the Dynkin diagram A3 with a linear orientation. By

applying the recursive construction to the initial seed (Q, x) one finds exactly
fourteen seeds (modulo simultaneous renumbering of vertices and cluster vari-
ables). These are the vertices of the exchange graph, which is isomorphic to
90 Bernhard Keller
the third Stasheff associahedron [122] [34]:

/.-,
()*+
o ooo 2%% RRRRRR
ooo %% RRR
oooo %
RRR
R
◦, %% ◦
,, %% ...
,, %
◦/
// ◦ OO /.-,
()*+ ◦
1 CC ◦
/ OOO
O {{{ C rrrr 444
◦> ◦C /.-,
()*+
>> CC {{ 3 q ◦
>> { qq
>> ◦
q qqqq
>> qq
>> qqqq
qq
◦
The vertex labeled 1 corresponds to (Q, x), the vertex 2 to μ2 (Q, x), which is
given by

& x1 + x3
1o 2o 3 , x1 , , x3 ,
x2
and the vertex 3 to μ1 (Q, x), which is given by

1o 2 / 3 , 1 + x2 , x2 , x3 .
x1
As expected (section 2.2), we find a total of 3 + 6 = 9 cluster variables, which
correspond bijectively to the faces of the exchange graph. The clusters x1 , x2 , x3
and x1 , x2 , x3 also appear naturally as slices of the repetition, where by a slice,
we mean a full connected subquiver containing a representative of each orbit
under the horizontal translation (a subquiver is full if, with any two vertices, it
contains all the arrows between them).
◦? ? x3 >> x3 x1
x
?? > ? >>
> ? >>> ? 3 ??
?
? >
x2 x2 x x x
~? ??? ? @@@ ? ???? ~~? 2
2
~ ? 2 @@@
~~
~
x1 x1 x1 x1 x3
In fact, as it is easy to check, each slice yields a cluster. However, some clusters
do not come from slices, for example the cluster x1 , x3 , x1 associated with the
seed μ2 (Q, x).
3.7. Cluster algebras with finitely many cluster variables

The phenomena observed in the above examples are explained by the following
key theorem:
Theorem 3.1 (Fomin-Zelevinsky [52]). Let Q be a finite connected quiver

without loops or 2-cycles with vertex set {1, . . . , n}. Let AQ be the associated
cluster algebra.
a) All cluster variables are Laurent polynomials, i.e. their denominators are
monomials.
b) The number of cluster variables is finite iff Q is mutation equivalent to an
orientation of a simply laced Dynkin diagram . In this case, is unique
and the non initial cluster variables are in bijection with the positive roots
of ; namely, if we denote the simple roots by α1 , . . . , αn , then for each

positive root d α , there is a unique non initial cluster variable whose
i i
denominator is xidi .
c) The knitting algorithm yields all cluster variables iff the quiver Q has two
vertices or is an orientation of a simply laced Dynkin diagram .
The theorem can be extended to the non simply laced case if we work with
valued quivers as in the example of G2 in section 2.2.
It is not hard to check that the knitting algorithm yields exactly the cluster
variables obtained by iterated mutations at sinks and sources. Remarkably, in
the Dynkin case, all cluster variables can be obtained in this way.
The construction of the cluster algebra shows that if the quiver Q is mutation-
equivalent to Q , then we have an isomorphism
∼
AQ → AQ
preserving clusters and cluster variables. Thus, to prove that the condition in b)
is sufficient, it suffices to show that AQ is cluster-finite if the underlying graph
of Q is a Dynkin diagram.
No normal form for mutation-equivalence is known in general and it is
unkown how to decide whether two given quivers are mutation-equivalent.
However, for certain restricted classes, the answer to this problem is known:
Trivially, two quivers with two vertices are mutation-equivalent iff they are
isomorphic. But it is already a non-trivial problem to decide when a quiver
1 ^<
r /2,
<<
s t
3
where r, s and t are non negative integers, is mutation-equivalent to a quiver
without a 3-cycle: As shown in [15], this is the case iff the ‘Markoff inequality’
r 2 + s 2 + t 2 − rst > 4
holds or one among r, s and t is < 2.

92 Bernhard Keller
For a general quiver Q, a criterion for AQ to be cluster-finite in terms of

quadratic forms was given in [14]. In practice, the quickest way to decide
whether a concretely given quiver is cluster-finite and to determine its cluster-
type is to compute its mutation-class using [87]. For example, the reader can
easily check that for 3 ≤ n ≤ 8, the following quiver glued together from n − 2
triangles
1o 3o 5o .? . . o n< − 1
@ @ ~~~ xx
~~ xx
~~ xx
~~ xxx
2o 4o 6o ... o n
is cluster-finite of respective cluster-type A3 , D4 , D5 , E6 , E7 and E8 and that
it is not cluster-finite if n > 8.
4. Cluster algebras with coefficients

In their combinatorial properties, cluster algebras with coefficients are very
similar to those without coefficients which we have considered up to now. The
great virtue of cluster algebras with coefficients is that they proliferate in nature
as algebras of coordinates on homogeneous varieties. We will define cluster
algebras with coefficients and illustrate their ubiquity on several examples.
4.1. Definition
with
Let 1 ≤ n ≤ m be integers. An ice quiver of type (n, m) is a quiver Q
vertex set
{1, . . . , m} = {1, . . . , n} ∪ {n + 1, . . . , m}
such that there are no arrows between any vertices i, j which are strictly greater
than n. The principal part of Q is the full subquiver Q of Q whose vertex set is
{1, . . . , n} (a subquiver is full if, with any two vertices, it contains all the arrows
between them). The vertices n + 1, . . . , m are often called frozen vertices. The
cluster algebra
AQ ⊂ Q(x1 , . . . , xm )
is defined as before but
– only mutations with respect to vertices in the principal part are allowed and
no arrows are drawn between the vertices > n,
– in a cluster
u = {u1 , . . . , un , xn+1 , . . . , xm }
only u1 , . . . , un are called cluster variables; the elements xn+1 , . . . , xm are
called coefficients; to make things clear, the set u is often called an extended
cluster;
– the cluster type of Q is that of Q if it is defined.
Often, one also considers localizations of AQ obtained by inverting certain

coefficients. Notice that the datum of Q corresponds to that of the integer

m × n-matrix B whose top n × n-submatrix B is antisymmetric and whose
entry bij equals the number of arrows i → j or the opposite of the number of
One can also
arrows j → i. The matrix B is called the principal part of B.
consider valued ice quivers, which will correspond to m × n-matrices whose
principal part is antisymmetrizable.
4.2. Example: SL(2, C)

Let us consider the algebra of regular functions on the algebraic group SL(2, C),
i.e. the algebra
C[a, b, c, d]/(ad − bc − 1).
We claim that this algebra has a cluster algebra structure, namely that it
is isomorphic to the complexification of the cluster algebra with coefficients
associated with the following ice quiver
1>
>>
>>
>>

2 3
where we have framed the frozen vertices. Indeed, here the principal part Q
only consists of the vertex 1 and we can only perform one mutation, whose
associated exchange relation reads
x1 x1 = 1 + x2 x3 or x1 x1 − x2 x3 = 1.
We obtain an isomorphism as required by sending x1 to a, x1 to d, x2 to b and
x3 to c. We describe this situation by saying that the quiver
a>
>>>
>>

>
b c
94 Bernhard Keller
whose vertices are labeled by the images of the corresponding variables, is

an initial seed for a cluster structure on the algebra A. Notice that this cluster
structure is not unique.
4.3. Example: Planes in affine space

As a second example, let us consider the algebra A of polynomial functions
on the cone over the Grassmannian of planes in Cn+3 . This algebra is the C-
algebra generated by the Plücker coordinates xij , 1 ≤ i < j ≤ n + 3, subject
to the Plücker relations: for each quadruple of integers i < j < k < l, we have
xik xj l = xij xkl + xj k xil .
Each plane P in Cn+3 gives rise to a straight line in this cone, namely the one
generated by the 2 × 2-minors xij of any (n + 3) × 2-matrix whose columns
generate P . Notice that the monomials in the Plücker relation are naturally
associated with the sides and the diagonals of the square
i << j
O << O
O< O
<< O
O
<
l k
The relation expresses the product of the variables associated with the diagonals
as the sum of the monomials associated with the two pairs of opposite sides.
Now the idea is that the Plücker relations are exactly the exchange relations
for a suitable structure of cluster algebra with coefficients on the coordinate
ring. To formulate this more precisely, let us consider a regular (n + 3)-gon in
the plane with vertices numbered 1, . . . , n + 2, and consider the variable xij as
associated with the segment [ij ] joining the vertices i and j .
Proposition 4.1 ([52, Example 12.6]). The algebra A has a structure of cluster
algebra with coefficients such that
– the coefficients are the variables xij associated with the sides of the (n + 3)-
gon;
– the cluster variables are the variables xij associated with the diagonals of
the (n + 3)-gon;
– the clusters are the n-tuples of variables whose associated diagonals form a
triangulation of the (n + 3)-gon.
Moreover, the exchange relations are exactly the Plücker relations and the
cluster type is An .
Thus, a triangulation of the (n + 3)-gon determines an initial seed for the

cluster algebra and hence an ice quiver Q whose frozen vertices correspond to
the sides of the (n + 3)-gon and whose non frozen variables to the diagonals
in the triangulation. The arrows of the quiver are determined by the exchange
relations which appear when we wish to replace one diagonal [ik] of the
triangulation by its flip, i.e. the unique diagonal [j l] different from [ik] which
does not cross any other diagonal of the triangulation. It is not hard to see that
this means that the underlying graph of Q is the graph dual to the triangulation
and that the orientation of the edges of this graph is induced by the choice
of an orientation of the plane. Here is an example of a triangulation and the
associated ice quiver:
o 0 OO
ooo 44 OO
ooo 44 OO
05 01
W//
5 44 1
44 45 / 04 gOOO 02 / 12
44 wooo O
4
03
? ???
4 NNN
NNN oooooo
2
o 34 23
3
4.4. Example: The big cell of the Grassmannian

We consider the cone over the big cell in the Grassmannian of k-dimensional
subspaces of the space of rows Cn , where 1 ≤ k ≤ n are fixed integers such that
l = n − k is greater or equal to 2. In more detail, let G be the group SL(n, C)
and P the subgroup of G formed by the block lower triangular matrices with
diagonal blocks of sizes k × k and l × l. The quotient P \ G identifies with
our Grassmannian. The big cell is the image under π : G → P \ G of the
space of block upper triangular matrices whose diagonal is the identity matrix
and whose upper right block is an arbitrary k × l-matrix Y . The projection π
induces an isomorphism between the space Mk×l (C) of these matrices and the
big cell. In particular, the algebra A of regular functions on the big cell is the
algebra of polynomials in the coefficients yij , 1 ≤ i ≤ k, 1 ≤ j ≤ l, of Y . Now
for 1 ≤ i ≤ k and 1 ≤ j ≤ l, let Fij be the largest square submatrix of Y whose
lower left corner is (i, j ) and let s(i, j ) be its size. Put
fij = (−1)(k−i)(s(i,j )−1) det(Fij ).
Theorem 4.2 ([69]). The algebra A has the structure of a cluster algebra with
coefficients whose initial seed is given by
96 Bernhard Keller
f11 / f12 / f13 ... / f1l

O | | O | | O O
|| ||
||| |||
~|| ~||
|
/ / f23 ... / f2l
f21 f22O
O | | O
| |
|| ||
|| ||
|| ||
|| ||

.. ~| .. ~| .. ..
.O | | .O . .O
|||
|
||
~||
/ fk2 / fk3 / fkl
fk1
The following table indicates when these algebras are cluster-finite and what
their cluster-type is:
k\n 2 3 4 5 6 7
2 A1 A2 A3 A4 A5 A6
3 A2 D4 E6 E8
4 A3 E6
5 A4 E8
The homogeneous coordinate ring of the Grassmannian Gr(k, n) itself also

has a cluster algebra structure [120] and so have partial flag varieties, double
Bruhat cells, Schubert varieties . . . , cf. [62] [17].
4.5. Compatible Poisson structures

Recall that the group SL(n, C) has a canonical Poisson structure given by the
Sklyanin bracket, which is defined by
ω(xij , xαβ ) = (sign(α − i) − sign(β − j ))xiβ xαj
where the xij are the coordinate functions on SL(n, C). This bracket makes
G = SL(n, C) into a Poisson-Lie group and P \ G into a Poisson G-variety for
each subgroup P of G containing the subgroup B of lower triangular matrices.
In particular, the big cell of the Grassmannian considered above inherits a
Poisson bracket.
Theorem 4.3 ([69]). This bracket is compatible with the cluster algebra struc-
ture in the sense that each extended cluster is a log-canonical coordinate system,
i.e. we have
ω(ui , uj ) = ωij(u) ui uj
for certain (integer) constants ωij(u) depending on the extended cluster u. More-
over, the coefficients are central for ω.
be an ice quiver.
This theorem admits the following generalization: Let Q
to be obtained by glueing the complex tori
Define the cluster variety X (Q)
indexed by the clusters u
T (u) = (C∗ )n = Spec(C[u1 , u−1 −1

1 , . . . , um , um ])
using the exchange relations as glueing maps, where m is the number of vertices

of Q.
Theorem 4.4 ([69]). Suppose that the principal part Q of Q is connected and

that the matrix B associated with Q is of maximal rank. Then the vector space
compatible with the cluster algebra structure is
of Poisson structures on X (Q)
of dimension

m−n
1+ .
2
is an open subset of the
Notice that in general, the cluster variety X (Q)
spectrum of the (complexified) cluster algebra. For example, for the cluster
algebra associated with SL(2, C) which we have considered above, the cluster
variety is the union of the elements

a b
c d
of SL(2, C) such that we have abc = 0 or bcd = 0. The cluster variety is
always regular, but the spectrum of the cluster algebra may be singular.
For example, the spectrum of the cluster algebra associated with the ice
quiver
x@
~~~ @@@ 2
~~ @@
2
~
~ @
u v
is the hypersurface in C4 defined by the equation xx = u2 + v 2 , which is

singular at the origin. The corresponding cluster variety is obtained by removing
the points with x = x = u2 + v 2 = 0 and is regular.
98 Bernhard Keller
In the above theorem, the assumption that B be of full rank is essential.

Otherwise, there may not exist any Poisson bracket compatible with the cluster
algebra structure. However, as shown in [70], for any cluster algebra with
coefficients, there are ‘dual Poisson structures’, namely certain 2-forms, which
are compatible with the cluster algebra structure.
4.6. Example: The maximal unipotent subgroup of SL(n + 1, C)

Let n be a non negative integer and N the subgroup of SL(n + 1, C) formed
by the upper triangular matrices with all diagonal coefficients equal to 1. For
1 ≤ i, j ≤ n + 1 and g ∈ N , let Fij (g) be the maximal square submatrix of g
whose lower left corner is (i, j ). Let fij (g) be the determinant of Fij (g). We
consider the functions fij for 1 ≤ i ≤ n and i + j ≤ n + 1.
Theorem 4.5 ([17]). The coordinate algebra C[N ] has an upper cluster alge-
bra structure whose initial seed is given by
f12 / f13 / f14 / . .O . / f1,n+1

O ~ O
~~

~~~
~~
~~
~ |
f22 / f23 / ... f2,n
.. / ...
.O
~
fn,2
We refer to [17] for the notion of ‘upper’ cluster algebra structure. It is not
hard to check that this structure is of cluster type A3 for n = 3, D6 for n = 4
and cluster-infinite for n ≥ 5. For n = 5, this cluster algebra is related to the
elliptic root system E8(1,1) in the notations of Saito [118], cf. [64].
A theorem of Fekete [42] generalized in [16] claims that a square matrix
of order n + 1 is totally positive (i.e. all its minors are > 0) if and only if the
following (n + 1)2 minors of g are positive: all minors occupying several initial
rows and several consecutive columns, and all minors occupying several initial
columns and several consecutive rows. It follows that an element g of N is
totally positive if fij (g) > 0 for the fij belonging to the initial seed above.
The same holds for the u1 , . . . , um in place of these fij for any cluster u of
this cluster algebra because each exchange relation expresses the new variable
subtraction-free in the old variables.
Geiss-Leclerc-Schröer have shown [65] that each monomial in the variables
of an arbitrary cluster belongs to Lusztig’s dual semicanonical basis of C[N ]
[104]. They also show that the dual semicanonical basis of C[N ] is different
from the dual canonical basis of Lusztig and Kashiwara except in types A2 , A3
and A4 [64].
5. Categorification via cluster categories: the finite case

5.1. Quiver representations and Gabriel’s theorem
We refer to the books [117] [59] [5] and [4] for a wealth of information on the
representation theory of quivers and finite-dimensional algebras. Here, we will
only need very basic notions.
Let Q be a finite quiver without oriented cycles. For example, Q can be an
orientation of a simply laced Dynkin diagram or the quiver
α ttt
9 2 KKK β
t KKK
tt %/
1 γ 3.
Let k be an algebraically closed field. A representation of Q is a diagram of

finite-dimensional vector spaces of the shape given by Q. More formally, a
representation of Q is the datum V of
r a finite-dimensional vector space Vi for each vertex i or Q,

r a linear map Vα : Vi → Vj for each arrow α : i → j of Q.
Thus, in the above example, a representation of Q is a (not necessarily com-

mutative) diagram
rr9
V2 LL V
r
Vα
r LLLβ
rr L%
V1 / V3
Vγ
formed by three finite-dimensional vector spaces and three linear maps. A

morphism of representations is a morphism of diagrams. More formally, a
morphism of representations f : V → W is the datum of a linear map fi :
100 Bernhard Keller
Vi → Wi for each vertex i of Q such that the square
Vi
Vα
/ Vj
fi fj

Wi / Wj
Wα
commutes for all arrows α : i → j of Q. The composition of morphisms is

defined in the natural way. We thus obtain the category of representations
rep(Q). A morphism f : V → W of this category is an isomorphism iff its
components fi are invertible for all vertices i of Q0 .
For example, let Q be the quiver
1 /2,
and
V : V1
Vα
/ V2
a representation of Q. By choosing basis in the spaces V1 and V2 we find an

isomorphism of representations
VO 1
Vα
/ V2
O
kn / kp ,
A
where, by abuse of notation, we denote by A the multiplication by a p × n-

matrix A. We know that we have

I 0
P AQ = r
0 0
for invertible matrices P and Q, where r is the rank of A. Let us denote the
right hand side by Ir ⊕ 0. Then we have an isomorphism of representations
kO n / kp
A O
Q P −1
kn / kp
Ir ⊕0
We thus obtain a normal form for the representations of this quiver.

Now the category repk (Q) is in fact an abelian category: Its direct sums,
kernels and cokernels are computed componentwise. Thus, if V and W are two
representations, then the direct sum V ⊕ W is the representation given by
(V ⊕ W )i = Vi ⊕ Wi and (V ⊕ W )α = Vα ⊕ Wα ,
for all vertices i and all arrows α of Q. For example, the above representation
in normal form is isomorphic to the direct sum
(k
1 / k )r ⊕ ( k / 0 )n−r ⊕ ( 0 / k )p−r .
The kernel of a morphism of representations f : V → W is given by
ker(f )i = ker(fi : Vi → Wi )
endowed with the maps induced by the Vα and similarly for the cokernel. A
subrepresentation V of a representation V is given by a family of subspaces
Vi ⊂ Vi , i ∈ Q0 , such that the image of Vi under Vα is contained in Vj for
each arrow α : i → j of Q. A sequence
0 /U /V /W /0
of representations is a short exact sequence if the sequence
0 / Ui / Vi / Wi /0
is exact for each vertex i of Q.

A representation V is simple if it is non zero and if for each subrepresentation
V of V we have V = 0 or V /V = 0. Equivalently, a representation is simple
if it has exactly two subrepresentations. A representation V is indecomposable
if it is non zero and in each decomposition V = V ⊕ V , we have V = 0 or
V = 0. Equivalently, a representation is indecomposable if it has exactly two
direct factors.
In the above example, the representations
k / 0 and 0 /k
are simple. The representation
V = (k
1 / k)
is not simple: It has the non trivial subrepresentation 0 / k . However, it

is indecomposable. Indeed, each endomorphism f : V → V is given by two
equal components f1 = f2 so that the endomorphism algebra of V is one-
dimensional. If V was a direct sum V ⊕ V for two non-zero subspaces, the
endomorphism algebra of V would contain the product of the endomorphism
algebras of V and V and thus would have to be at least of dimension 2. Since
102 Bernhard Keller
V is indecomposable, the exact sequence
0 → (0 / k) → (k 1 / k) → (k / 0) → 0
is not a split exact sequence.

If Q is an arbitrary quiver, for each vertex i, we define the representation Si
by

k i=j
(Si )j =
0 else.
Then clearly the representations Si are simple and pairwise non isomorphic. As
an exercise, the reader may show that if Q does not have oriented cycles, then
each representation admits a finite filtration whose subquotients are among the
Si . Thus, in this case, each simple representation is isomorphic to one of the
representations Si .
Recall that a (possibly non commutative) ring is local if its non invertible
elements form an ideal.
Decomposition Theorem 5.1 (Azumaya-Fitting-Krull-Remak-Schmidt).
a) A representation is indecomposable iff its endomorphism algebra is local.

b) Each representation decomposes into a finite sum of indecomposable rep-
resentations, unique up to isomorphism and permutation.
As we have seen above, for quivers without oriented cycles, the classifica-
tion of the simple representations is trivial. On the other hand, the problem of
classifying the indecomposable representations is non trivial. Let us examine
this problem in a few examples: For the quiver 1 → 2, we have checked the
existence in part b) directly. The uniqueness in b) then implies that each inde-
composable representation is isomorphic to exactly one of the representations
S1 , S2 and
k
1 / k.
Similarly, using elementary linear algebra it is not hard to check that each
indecomposable representation of the quiver
An : 1 /2 / ... /n
is isomorphic to a representation I [p, q], 1 ≤ p < q ≤ n, which takes the

vertices i in the interval [p, q] to k, the arrows linking them to the identity and
all other vertices to zero. In particular, the number of isomorphism classes of
indecomposable representations of An is n(n + 1)/2.
The representations of the quiver
1d α
are the pairs (V1 , Vα ) consisting of a finite-dimensional vector space and an

endomorphism and the morphisms of representations are the ‘intertwining oper-
ators’. It follows from the existence and uniqueness of the Jordan normal form
that a system of representatives of the isomorphism classes of indecomposable
representations is formed by the representations (k n , Jn,λ ), where n ≥ 1 is an
integer, λ a scalar and Jn,λ the Jordan block of size n with eigenvalue λ.
The Kronecker quiver
/
1 /2
admits the following infinite family of pairwise non isomorphic representations:
λ /
k /k,
μ
where (λ : μ) runs through the projective line.
Question 5.2. For which quivers are there only finitely many isomorphism
classes of indecomposable representations?
To answer this question, we define the dimension vector of a representation

V to be the sequence dim V of the dimensions dim Vi , i ∈ Q0 . For example, the
dimension vectors of the indecomposable representations of A2 are the pairs
dim S1 = [10] , dim S2 = [01] , dim (k → k) = [11].
We define the Tits form
qQ : ZQ0 → Z
by

qQ (v) = vi2 − vs(α) vt(α) .
i∈Q0 α∈Q1
Notice that the Tits form does not depend on the orientation of the arrows of Q
but only on its underlying graph. We say that the quiver Q is representation-
finite if, up to isomorphism, it has only finitely many indecomposable repre-
sentations. We say that a vector v ∈ ZQ0 is a root of qQ if qQ (v) = 1 and that
it is positive if its components are ≥ 0.
Theorem 5.3 (Gabriel [58]). Let Q be a connected quiver and assume that k
is algebraically closed. The following are equivalent:
104 Bernhard Keller
(i) Q is representation-finite;
(ii) qQ is positive definite;
(iii) The underlying graph of Q is a simply laced Dynkin diagram .
Moreover, in this case, the map taking a representation to its dimension vec-
tor yields a bijection from the set of isomorphism classes of indecomposable
representations to the set of positive roots of the Tits form qQ .
It is not hard to check that if the conditions hold, the positive roots of qQ are
in turn in bijection with the positive roots of the root system associated with
, via the map taking a positive root v of qQ to the element

vi αi
i∈Q0
of the root lattice of .

Let us consider the example of the quiver Q = A2 . In this case, the Tits
form is given by
qQ (v) = v12 + v22 − v1 v2 .
It is positive definite and its positive roots are indeed precisely the dimension
vectors
[01] , [10] , [11]
of the indecomposable representations.

Gabriel’s theorem has been generalized to non algebraically closed ground
fields by Dlab and Ringel [40]. Let us illustrate the main idea on one simple
example: Consider the category of diagrams
V : V1
f
/ V2
where V1 is a finite-dimensional real vector space, V2 a finite-dimensional

complex vector space and f an R-linear map. Morphisms are given in the
natural way. Then we have the following complete list of representatives of the
isomorphism classes of indecomposables:
R → 0 , R2 → C , R → C , 0 → C.
The corresponding dimension vectors are
[10] , [21] , [11] , [01].
They correspond bijectively to the positive roots of the root system B2 .

5.2. Tame and wild quivers

The quivers with infinitely many isomorphism classes of indecomposables
can be further subdivided into two important classes: A quiver is tame if it
has infinitely many isomorphism classes of indecomposables but these occur
in ‘families of at most one parameter’ (we refer to [117] [4] for the precise
definition). The Kronecker quiver is a typical example. A quiver is wild if
there are ‘families of indecomposables of ≥ 2 parameters’. One can show that
in this case, there are families of an arbitrary number of parameters and that
the classification of the indecomposables over any fixed wild algebra would
entail the classification of the the indecomposables over all finite-dimensional
algebras. The following three quivers are representation-finite, tame and wild
respectively:
1= 1-< 1-=
=== ---<<< ---===
== -- << -- ==

= < =
2 3 4 2 3 4 5 2 3 4 5 6.
Theorem 5.4 (Donovan-Freislich [41], Nazarova [110]). Let Q be a connected

quiver and assume that k is algebraically closed. Then Q is tame iff the under-
lying graph of Q is a simply laced extended Dynkin diagram.
Let us recall the list of simply laced extended Dynkin quivers. In each case,
the number of vertices of the diagram D n equals n + 1.
n : ◦ SSSS
A kkk SSS
kkkk . . .
◦ ◦ ◦ ◦
n :
D ◦ EE ◦
E yyy
◦ ◦ ... ◦ ◦ EE
yyy E
◦ ◦
6 :
E ◦ ◦ ◦ ◦ ◦
◦
◦
7 :
E ◦ ◦ ◦ ◦ ◦ ◦ ◦
◦
8 : ◦
E ◦ ◦ ◦ ◦ ◦ ◦ ◦
◦
The following theorem is a first illustration of the close connection between

cluster algebras and the representation theory of quivers. Let Q be a finite quiver
106 Bernhard Keller
without oriented cycles and let ν(Q) be the supremum of the multiplicities of
the arrows occurring in all quivers mutation-equivalent to Q.
Theorem 5.5.
a) Q is representation-finite iff ν(Q) equals 1.
b) Q is tame iff ν(Q) equals 2.
c) Q is wild iff ν(Q) ≥ 3 iff ν(Q) = ∞.
d) The mutation class of Q is finite iff Q has two vertices, is representation-
finite or tame.
Here, part a) follows from Gabriel’s theorem and part (iii) of Theorem 1.8
in [52]. Part b) follows from parts a) and c) by exclusion of the third. For part
c), let us first assume that Q is wild. Then it is proved at the end of the proof of
theorem 3.1 in [13] that ν(Q) = ∞. Conversely, let us assume that ν(Q) ≥ 3.
Then using Theorem 5 of [29] we obtain that Q is wild. Part d) is proved in
[13].
5.3. The Caldero-Chapoton formula

Let be a simply laced Dynkin diagram and Q a quiver with underlying graph
. Suppose that the set of vertices of and Q is the set of the natural numbers
1, 2, . . . , n. We already know from part b) of theorem 3.1 that for each positive
root

n
α= di αi
i=1
of the corresponding root system, there is a unique non initial cluster variable
Xα with denominator
x1d1 . . . xndn .
By combining this with Gabriel’s theorem, we get the
Corollary 5.6. The map taking an indecomposable representation V with
dimension vector (di ) of Q to the unique non initial cluster variable XV whose
denominator is x1d1 . . . xndn induces a bijection from the set of ismorphism classes
of indecomposable representations to the set of non initial cluster variables.
Let us consider this bijection for Q = A2 :
S2 = (0 → k) P1 = (k → k) S1 = (k → 0)
1 + x1 x1 + 1 + x2 1 + x2
XS2 = XP1 = SS1 =
x2 x1 x2 x1
We observe that for the two simple representations, the numerator contains
exactly two terms: the number of subrepresentations of the simple represen-
tation! Moreover, the representation P1 has exactly three subrepresentations
and the numerator of XP1 contains three terms. In fact, it turns out that this
phenomenon is general in type A. But now let us consider the following quiver
with underlying graph D4
3 NNN
NNN
'/
2
ppp7 4
ppp
1
and the dimension vector d with d1 = d2 = d3 = 1 and d4 = 2. The unique

(up to isomorphism) indecomposable representation V with dimension vector
d consists of a plane V4 together with three lines in general position Vi ⊂ V4 ,
i = 1, 2, 3. The corresponding cluster variable is
1
X4 = (1 + 3x4 + 3x42 + x43 + 2x1 x2 x3 + 3x1 x2 x3 x4 + x12 x22 x32 ).
x1 x2 x3 x42
Its numerator contains a total of 14 monomials. On the other hand, it is easy

to see that V4 has only 13 types of submodules: twelve submodules are deter-
mined by their dimension vectors but for the dimension vector e = (0, 0, 0, 1),
we have a family of submodules: Each submodule of this dimension vector
corresponds to the choice of a line in V4 . Thus for this dimension vector e,
the family of submodules is parametrized by a projective line. Notice that the
Euler characteristic of the projective line is 2 (since it is a sphere: the Riemann
sphere). So if we attribute weight 1 to the submodules determined by their
dimension vector and weight 2 to this P1 -family, we find a ‘total submodule
weight’ equal to the number of monomials in the numerator. These considera-
tions led Caldero-Chapoton [27] to the following definition, whose ingredients
we describe below: Let Q be a finite quiver with vertices 1, . . . , n, and V a
finite-dimensional representation of Q. Let d be the dimension vector of V .
Define
! "
1 n
j →i ej + i→j (dj −ej )
CC(V ) = d1 d2 χ (Gre (V )) xi .
x1 x2 . . . xndn 0≤e≤d i=1
Here the sum is taken over all vectors e ∈ Nn such that 0 ≤ ei ≤ di for all i. For
each such vector e, the quiver Grassmannian Gre (V ) is the variety of n-tuples of
subspaces Ui ⊂ Vi such that dim Ui = ei and the Ui form a subrepresentation
of V . By taking such a subrepresentation to the family of the Ui , we obtain a
108 Bernhard Keller
map

n
Gre (V ) → Grei (Vi ) ,
i=1
where Grei (Vi ) denotes the ordinary Grassmannian of ei -dimensional subspaces

of Vi . Recall that the Grassmannian carries a canonical structure of projective
variety. It is not hard to see that for a family of subspaces (Ui ) the condition of
being a subrepresentation is a closed condition so that the quiver Grassmannian
identifies with a projective subvariety of the product of ordinary Grassmannians.
If k is the field of complex numbers, the Euler characteristic χ is taken with
respect to singular cohomology with coefficients in Q (or any other field). If k
is an arbitrary algebraically closed field, we use étale cohomology to define χ .
The most important properties of χ are (cf. e.g. section 7.4 in [67])
(1) χ is additive with respect to disjoint unions;

(2) if p : E → X is a morphism of algebraic varieties such that the Euler
characteristic of the fiber over a point x ∈ X does not depend on x, then
χ (E) is the product of χ (X) by the Euler characteristic of the fiber over
any point x ∈ X.
Theorem 5.7 (Caldero-Chapoton [27]). Let Q be a Dynkin quiver and V

an indecomposable representation. Then we have CC(V ) = XV , the cluster
variable obtained from V by composing Fomin-Zelevinsky’s bijection with
Gabriel’s.
Caldero-Chapoton’s proof of the theorem was by induction. One of the aims

of the following sections is to explain ‘on what’ they did the induction.
5.4. The derived category

Let k be an algebraically closed field and Q a (possibly infinite) quiver without
oriented cycles (we will impose more restrictive conditions on Q later). For
example, Q could be the quiver
/3
1
@
α
γ

β
2 /4
A path of Q is a formal composition of ≥ 0 arrows. For example, the sequence

(4|α|β|γ |1) is a path of length 3 in the above example (notice that we include
the source and target vertices of the path in the notation). For each vertex i of
Q, we have the lazy path ei = (i|i), the unique path of length 0 which starts
at i and stops at i and does nothing in between. The path category has set of
objects Q0 (the set of vertices of Q) and, for any vertices i, j , the morphism
space from i to j is the vector space whose basis consists of all paths from i
to j . Composition is induced by composition of paths and the unit morphisms
are the lazy paths. If Q is finite, we define the path algebra to be the matrix
algebra

kQ = Hom(i, j )
i,j ∈Q0
where multiplication is matrix multiplication. Equivalently, the path algebra

has as a basis all paths and its product is given by concatenating composable
paths and equating the product of non composable paths to zero. The path
algebra has the sum of the lazy paths as its unit element

1= ei .
i∈Q0
The idempotent ei yields the projective right module
Pi = ei kQ.
The modules Pi generate the category of k-finite-dimensional right modules

mod kQ. Each arrow α from i to j yields a map Pi → Pj given by left mul-
tiplication by α. (If we were to consider – heaven forbid – left modules, the
analogous map would be given by right multiplication by α and it would go in
the direction opposite to that of α. Whence our preference for right modules).
Notice that we have an equivalence of categories
repk (Qop ) → mod kQ
sending a representation V of the opposite quiver Qop to the sum

Vi
i∈Q0
endowed with the natural right action of the path algebra. Conversely, a kQ-
module M gives rise to the representation V with Vi = Mei for each vertex i
of Q and Vα given by right multiplication by α for each arrow α of Q. The
category mod kQ is abelian, i.e. it is additive, has kernels and cokernels and
for each morphism f the cokernel of its kernel is canonically isomorphic to the
kernel of its cokernel.
The category mod kQ is hereditary. Recall from [32] that this means that
submodules of projective modules are projective; equivalently, that all extension
110 Bernhard Keller
groups in degrees i ≥ 2 vanish:

ExtikQ (L, M) = 0;
equivalently, that kQ is of global dimension ≤ 1; . . . Thus, in the spirit of

noncommutative algebraic geometry approached via abelian categories, we
should think of mod kQ as a ‘non commutative curve’.
We define DQ to be the bounded derived category Db (mod kQ) of the
abelian category mod kQ. Thus, the objects of DQ are the bounded complexes
of (right) kQ-modules
dp
. . . → 0 → . . . → M p → M p+1 → . . . → 0 → . . . .
Its morphisms are obtained from morphisms of complexes by formally inverting
all quasi-isomorphisms. We refer to [125] [84] . . . for in depth treatments of
the fundamentals of this construction. Below, we will give a complete and
elementary description of the category DQ if Q is a Dynkin quiver. We have
the following general facts: The functor
mod kQ → DQ
taking a module M to the complex concentrated in degree 0

... → 0 → M → 0 → ...
is a fully faithful embedding. From now on, we will identify modules with
complexes concentrated in degree 0. If L and M are two modules, then we
have a canonical isomorphism
∼
ExtikQ (L, M) → HomDQ (L, M[i])
for all i ∈ Z, where M[i] denotes the complex M shifted by i degrees to the left:
M[i]p = M p+i , p ∈ Z, and endowed with the differential dM[i] = (−1)i dM .
The category DQ has all finite direct sums (and they are given by direct sums of
complexes) and the decomposition theorem 5.1 holds. Moreover, each object
is isomorphic to a direct sum of shifted copies of modules (this holds more
generally in the derived category of any hereditary abelian category, for example
the derived category of coherent sheaves on an algebraic curve). The category
DQ is abelian if and only if the quiver Q does not have any arrows. However,
it is always triangulated. This means that it is k-linear (it is additive, and
the morphism sets are endowed with k-vector space structures so that the
composition is bilinear) and endowed with the following extra structure:
a) a suspension (or shift) functor : DQ → DQ , namely the functor taking a
complex M to M[1];
b) a class of triangles (sometimes called ‘distinguished triangles’), namely the

sequences
L → M → N → L
which are ‘induced’ by short exact sequences of complexes.
The class of triangles satisfies certain axioms, cf. e.g. [125]. The most important
consequence of these axioms is that the triangles induce long exact sequences
in the functors Hom(X, ?) and Hom(?, X), i.e. for each object X of DQ , the
sequences
. . . (X, −1 N ) → (X, L) → (X, M) → (X, N ) → (X, L) → . . .
and
. . . ( −1 N, X) ← (L, X) ← (M, X) ← (N, X) ← (L, X) ← . . .
are exact.
5.5. Presentation of the derived category of a Dynkin quiver

From now on, we assume that Q is a Dynkin quiver. Let ZQ be its repetition
(cf. section 2.2). So the vertices of ZQ are the pairs (p, i), where p is an integer
and i a vertex of Q and the arrows of ZQ are obtained as follows: each arrow
α : i → j of Q yields the arrows
(p, α) : (p, i) → (p, j ) , p ∈ Z ,
and the arrows
σ (p, α) : (p − 1, j ) → (p, i) , p ∈ Z.
We extend σ to a map defined on all arrows of ZQ by defining
σ (σ (p, α)) = (p − 1, α).
We endow ZQ with the map σ and with the automorphism τ : ZQ → ZQ
taking (p, i) to (p − 1, i) and (p, α) to (p − 1, α) for all vertices i of Q, all
arrows α of Q and all integers p.
For a vertex v of ZQ the mesh ending at v is the full subquiver
= u1 B (5.5.1)
σ (α) zzz u BBBα
z
zz oo7
2 NN B
NNNBB
zozooo . N'
τ v OO 7v
OOO .. pppp
O' pp
us
112 Bernhard Keller
ind DA5
(1,5) • •
(1,4) • • ΣP2 • •
(1,3) • P3 • • • • •
(1,2) • P2 • • • •Σ 2
P2 • •
(1,1) • (2,1) • • • • • • •
Figure 1. The repetition of type An
formed by v, τ (v) and all sources u of arrows α : u → v of ZQ ending in v.

We define the mesh ideal to be the (two-sided) ideal of the path category of ZQ
which is generated by all mesh relators

rv = ασ (α) ,
arrows α:u→v
where v runs through the vertices of ZQ. The mesh category is the quotient of
the path category of ZQ by the mesh ideal.
Theorem 5.8 (Happel [73]).

a) There is a canonical bijection v
→ Mv from the set of vertices of ZQ to
the set of isomorphism classes of indecomposables of DQ which takes the
vertex (1, i) to the indecomposable projective Pi .
b) Let ind DQ be the full subcategory of indecomposables of DQ. The bijection
of a) lifts to an equivalence of categories from the mesh category of ZQ to
the category ind DQ .
In figure 1, we see the repetition for Q = A5 and the map taking its vertices
to the indecomposable objects of the derived category. The vertices marked •
belonging to the left triangle are mapped to indecomposable modules. The
vertex (1, i) corresponds to the indecomposable projective Pi . The arrow
(1, i) → (1, i + 1), 1 ≤ i ≤ 5, is mapped to the left multiplication by the arrow
i → i + 1. The functor takes a mesh (5.5.1) to a triangle
Mτ v / s Mu i / Mv / Mτ v (5.5.2)
i=1
called an Auslander-Reiten triangle or almost split triangle, cf. [74]. If Mv and

Mτ v are modules, then so is the middle term and the triangle comes from an
exact sequence of modules
0 / Mτ v / s Mu i / Mv /0
i=1
11111 −10000 −01000 −00100 −00010 −00001
11110 01111 −11000 −01100 −00110 −00011
11100 01110 00111 −11100 −01110 −00111
11000 01100 00110 00011 −11110 −01111
10000 01000 00100 00010 00001 −11111
Figure 2. Some dimension vectors of indecomposables in DA5
called an Auslander-Reiten sequence or almost split sequence, cf. [5]. These

almost split triangles respectively sequences can be characterized intrinsically
in DQ respectively mod kQ.
Recall that the Grothendieck group K0 (T ) of a triangulated category is the
quotient of the free abelian group on the isomorphism classes [X] of objects X
of T by the subgroup generated by all elements
[X] − [Y ] + [Z]
arising from triangles (X, Y, Z) of T . In the case of DQ , the natural map
K0 (mod kQ) → K0 (DQ )
is an isomorphism (its inverse sends a complex to the alternating sum of

the classes of its homologies). Since K0 (mod kQ) is free on the classes [Si ]
associated with the simple modules, the same holds for K0 (DQ ) so that its
elements are given by n-tuples of integers. We write dim M for the image in
K0 (DQ ) of an object M of K0 (DQ ) and call dim M the dimension vector of M.
Then each triangle (5.5.2) yields an equality

s
dim Mv = dim Mui − dim Mτ v .
i=1
Using these equalities, we can easily determine dim M for each indecomposable
M starting from the known dimension vectors dim Pi , 1 ≤ i ≤ n. In the above
example, we find the dimension vectors listed in figure (2).
Thanks to the theorem, the automorphism τ of the repetition yields a k-
linear automorphism, still denoted by τ , of the derived category DQ . This
automorphism has several intrinsic descriptions:
1) As shown in [60], it is the right derived functor of the left exact Coxeter
functor rep(Qop ) → rep(Qop ) introduced by Bernstein-Gelfand-Ponomarev
[19] in their proof of Gabriel’s theorem. If we identify K0 (DQ ) with the root
114 Bernhard Keller
lattice via Gabriel’s theorem, then the automorphism induced by τ −1 equals the
the Coxeter transformation c. As shown by Gabriel [60], the identity ch = 1,
where h is the Coxeter number, lifts to an isomorphism of functors
τ −h → 2 .
∼
(5.5.3)
2) It can be expressed in terms of the Serre functor of DQ : Recall that for
a k-linear triangulated category T with finite-dimensional morphism spaces,
a Serre functor is an autoequivalence S : T → T such that the Serre duality
formula holds: We have bifunctorial isomorphisms
∼
D Hom(X, Y ) → Hom(Y, SX) , X, Y ∈ T ,
where D is the duality Homk (?, k) over the ground field. Notice that this
determines the functor S uniquely up to isomorphism. In the case of DQ =
Db (mod kQ), it is not hard to prove that a Serre functor exists (it is given by
the left derived functor of the tensor product by the bimodule D(kQ)). Now
the autoequivalence τ , the suspension functor and the Serre functor S are
linked by the fundamental isomorphism
∼
τ → S. (5.5.4)
5.6. Caldero-Chapoton’s proof

The above description of the derived category yields in particular a description
of the module category, which is a full subcategory of the derived category. This
description was used by Caldero-Chapoton [27] to prove their formula. Let us
sketch the main steps in their proof: Recall that we have defined a surjective
map v
→ Xv from the set of vertices of the repetition to the set of cluster
variables such that
a) we have X(0,i) = xi for 1 ≤ i ≤ n and
b) we have

Xτ v Xv = 1 + Xw
arrows w→v
for all vertices v of the repetition.

We wish to show that we have
Xv = CC(Mv )
for all vertices v such that Mv is an indecomposable module. This is done by
induction on the distance of v from the vertices (1, i) in the quiver ZQ. More
precisely, one shows the following
a) We have CC(Pi ) = X(1,i) for each indecomposable projective Pi . Here

we use the fact that submodules of projectives are projective in order to
explicitly compute CC(Pi ).
b1) For each split exact sequence
0→L→E →M →0,
we have
CC(L)CC(M) = CC(E).
Thus, if E = E1 ⊕ . . . Es is a decomposition into indecomposables, then

s
CC(E) = CC(Ei ).
i=1
b2) If
0→L→E→M→0
is an almost split exact sequence, then we have
CC(E) + 1 = CC(L)CC(M).
It is now clear how to prove the equality Xv = CC(Mv ) by induction by

proceeding from the projective indecomposables to the right.
5.7. The cluster category

The cluster category
CQ = DQ /(τ −1 )Z = DQ /(S −1 2 )Z
is the orbit category of the derived category under the action of the cyclic group
generated by the autoequivalence τ −1 = S −1 2 . This means that the objects
of CQ are the same as those of the derived category DQ and that for two objects
X and Y , the morphism space from X to Y in CQ is

CQ (X, Y ) = DQ (X, (S −1 2 )p Y ).
p∈Z
Morphisms are composed in the natural way. This definition is due to Buan-
Marsh-Reineke-Reiten-Todorov [10], who were trying to obtain a better under-
standing of the ‘decorated quiver representations’ introduced by Reineke-
Marsh-Zelevinsky [106]. For quivers of type A, an equivalent category was
defined independently by Caldero-Chapoton-Schiffler [28] using an entirely
116 Bernhard Keller
different description. Clearly the category CQ is k-linear. It is not hard to check

that its morphism spaces are finite-dimensional.
One can show [90] that CQ admits a canonical structure of triangulated
category such that the projection functor π : DQ → CQ becomes a triangle
functor (in general, orbit categories of triangulated categories are no longer
triangulated). The Serre functor S of DQ clearly induces a Serre functor in CQ ,
which we still denote by S. Now, by the definition of CQ (and its triangulated
structure), we have an isomorphism of triangle functors
∼
S → 2.
This means that CQ is 2-Calabi-Yau. Indeed, for an integer d ∈ Z, a triangulated

category T with finite-dimensional morphism spaces is d-Calabi-Yau if it
admits a Serre functor isomorphic as a triangle functor to the dth power of its
suspension functor.
5.8. From cluster categories to cluster algebras

We keep the notations and hypotheses of the previous section. The suspension
functor and the Serre functor S induce automorphisms of the repetition ZQ
which we still denote by and S respectively. The orbit quiver ZQ/(τ −1 )Z
inherits the automorphism τ and the map σ (defined on arrows only) and
thus has a well-defined mesh category. Recall that we write Ext1 (X, Y ) for
Hom(X, Y ) in any triangulated category.
Theorem 5.9 ([10] [11]).

a) The decomposition theorem holds for the cluster category and the mesh
category of ZQ/(τ −1 )Z is canonically equivalent to the full subcategory
ind CQ of the indecomposables of CQ . Thus, we have an induced bijection
L
→ XL from the set of isomorphism classes of indecomposables of CQ to
the set of all cluster variables of AQ which takes the shifted projective Pi
to the initial variable xi , 1 ≤ i ≤ n.
b) Under this bijection, the clusters correspond to the cluster-tilting sets, i.e.
the sets of pairwise non isomorphic indecomposables T1 , . . . , Tn such that
we have
Ext 1 (Ti , Tj ) = 0
for all i, j .
c) If T1 , . . . , Tn is cluster-tilting, then the quiver (cf. below) of the endomor-

phism algebra of the sum T = ni=1 Ti does not have loops nor 2-cycles and
the associated antisymmetric matrix is the exchange matrix of the unique

seed containing the cluster XT1 , . . . , XTn .
In part b), the condition implies in particular that Ext1 (Ti , Ti ) vanishes. How-
ever, for a Dynkin quiver Q, we have Ext1 (L, L) = 0 for each indecomposable
L of CQ . A cluster-tilting object of CQ is the direct sum of the objects T1 , . . . , Tn
of a cluster-tilting set. Since these are pairwise non-isomorphic indecompos-
ables, the datum of T is equivalent to that of the Ti . A cluster-tilted algebra of
type Q is the endomorphism algebra of a cluster-tilting object of CQ . In part c),
the most subtle point is that the quiver does not have loops or 2-cycles [11]. Let
us recall what one means by the quiver of a finite-dimensional algebra over an
algebraically closed field:
Proposition-Definition 5.10 (Gabriel). Let B be a finite-dimensional algebra
over the algebraically closed ground field k.
a) There exists a quiver QB , unique up to isomorphism, such that B is Morita
equivalent to the algebra kQB /I , where I is an ideal of kQB contained in
the square of the ideal generated by the arrows of QB .
b) The ideal I is not unique in general but we have I = 0 iff B is hereditary.
c) There is a bijection i
→ Si between the vertices of QB and the isomorphism
classes of simple B-modules. The number of arrows from a vertex i to a
vertex j equals the dimension of Ext1B (Sj , Si ).
In our case, the algebra B is the endomorphism algebra of the sum T of the
cluster-tilting set T1 , . . . , Tn in CQ . In this case, the Morita equivalence of a)
even becomes an isomorphism (because the Ti are pairwise non isomorphic).
For a suitable choice of this isomorphism, the idempotent ei associated with
the vertex i is sent to the identity of Ti and the images of the arrows from i to
j yield a basis of the space of irreducible morphisms
irrT (Ti , Tj ) = radT (Ti , Tj )/ rad2T (Ti , Tj ) ,
where radT (Ti , Tj ) denotes the vector space of non isomorphisms from Ti to
Tj (thanks to the locality of the endomorphism rings, this set is indeed closed
under addition) and rad2T the subspace of non isomorphisms admitting a non
trivial factorization:

n
rad2T (Ti , Tj ) = radT (Tr , Tj ) radT (Ti , Tr ).
r=1
As an illustration of theorem 5.9, we consider the cluster-tilting set

T1 , . . . , T5 in CA5 depicted in figure 3. Here the vertices labeled 0, 1, . . . ,
4 have to be identified with the vertices labeled 20, 21, . . . , 24 (in this order) to
118 Bernhard Keller
T3
? ???
4 20
? ???? ? ???
? ?
?3? ?8? ? 13 ? ? 21 ?
??? ??? ??? ???

? 2 ?? T
? ??2 12
? ? 16
? ? ? 22 ?
?? ?? ??? ??? ???

? 1 ?? ? 6 ?? ?11 ?? ? 15 ? ? 18 ?
? ? 23 ?
?? ?? ?? ??? ?? ???

0 T1 10 T4 17 T5 24
Figure 3. A cluster-tilting set in A5
obtain the orbit quiver ZQ/(τ −1 )Z . In the orbit category, we have τ → so
∼
that T1 is the indecomposable associated to vertex 0, for example. Using this
and the description of the morphisms in the mesh category, it is easy to check
that we do have
Ext1 (Ti , Tj ) = 0
for all i, j . It is also easy to determine the spaces of morphisms
HomCQ (Ti , Tj )
and the compositions of morphisms. Determining these is equivalent to deter-

mining the endomorphism algebra

End(T ) = Hom(T , T ) = Hom(Ti , Tj ).
i,j
This algebra is easily seen to be isomorphic to the algebra given by the following
quiver Q
5 =o
γ
3
== @
== α
β
@ 2 == β
α
==
=
1 o 4
γ
with the relations
αβ = 0 , βγ = 0 , γ α = 0.
Thus the quiver of End(T ) is Q . It encodes the exchange matrix of the associ-
ated cluster
1 + x2
XT1 =
x1
x1 x2 + x1 x4 + x3 x4 + x2 x3 x4
XT2 =
x1 x2 x3
x1 x2 x3 + x1 x2 x3 x4 + x1 x2 x5 + x1 x4 x5 + x3 x4 x5 + x2 x3 x4 x5
XT3 =
x1 x2 x3 x4 x5
x2 + x4
XT4 =
x3
1 + x4
XT5 = .
x5
5.9. A K-theoretic interpretation of the exchange matrix

Keep the notations and hypotheses of the preceding section. Let T1 , . . . , Tn be
a cluster-tilting set, T the sum of the Ti and B its endomorphism algebra. For
two finite-dimensional right B-modules L and M put
L, Ma = dim Hom(L, M) − dim Ext1 (L, M)

− dim Hom(M, L) + dim Ext1 (M, L).
This is the antisymmetrization of a truncated Euler form. A priori it is defined

on the split Grothendieck group of the category mod B (i.e. the quotient of
the free abelian group on the isomorphism classes divided by the subgroup
generated by all relations obtained from direct sums in mod B).
Proposition 5.11 (Palu). The form , a descends to an antisymmetric form

on K0 (mod B). Its matrix in the basis of the simples is the exchange matrix
associated with the cluster corresponding to T1 , . . . , Tn .
5.10. Mutation of cluster-tilting sets

Let us recall two axioms of triangulated categories:
TR1 For each morphism u : X → Y , there exists a triangle

u
X → Y → Z → X.
TR2 A sequence
u v w
X → Y → Z → X
120 Bernhard Keller
is a triangle if and only if the sequence

v w −u
Y → Z → X → Y
is a triangle.
One can show that in TR1, the triangle is unique up to (non unique) isomor-
phism. In particular, up to isomorphism, the object Z is uniquely determined by
u. Notice the sign in TR2. It follows from TR1 and TR2 that a given morphism
also occurs as the second (respectively third) morphism in a triangle.
Now, with the notations and hypotheses of the preceding section, suppose
that T1 , . . . , Tn is a cluster-tilting set and Q the quiver of the endomorphism
algebra B of the sum of the Ti . As explained after proposition-definition 5.10,
we have a surjective algebra morphism

kQ → Hom(Ti , Tj )
i,j
which takes the idempotent ei to the identity of Ti and the arrows i → j to

irreducible morphisms Ti → Tj , for all vertices i, j of Q (cf. the above example
computation of B and Q = QB ).
Now let k be a vertex of Q (the mutating vertex). We choose triangles
u
Tk → Ti → Tk∗ → Tk
arrows
k→i
and
v
∗
Tk → Tj → Tk → ∗Tk ,
arrows
j →k
where the component of u (respectively v) corresponding to an arrow α : k → i

(respectively j → k) is the corresponding morphism Tk → Ti (respectively
Tj → Tk ). These triangles are unique up to isomorphism and called the
exchange triangles associated with k and T1 , . . . , Tn .
Theorem 5.12 ([10]).
a) The objects Tk∗ and ∗Tk are isomorphic.
b) The set obtained from T1 , . . . , Tn by replacing Tk with Tk∗ is cluster-tilting
and its associated cluster is the mutation at k of the cluster associated with
T 1 , . . . , Tn .
c) Two indecomposables L and M appear as the the pair (Tk , Tk∗ ) associated
with an exchange if and only if the space Ext1 (L, M) is one-dimensional.
In this case, the exchange triangles are the unique (up to isomorphism) non
split triangles
L → E → M → L and M → E → L → M.
Let us extend the map L

→ XL from indecomposable to decomposable
objects of CQ by requiring that we have
XN = XN1 XN2
whenever N = N1 ⊕ N2 (this is compatible with the muliplicativity of the

Caldero-Chapoton map). We know that if u1 , . . . , un is a cluster and B = (bij )
the associated exchange matrix, then the mutation at k yields the variable uk
such that

uk uk = ui + uj .
arrows arrows
k→i j →k
By combining this with the exchange triangles, we see that in the situation of
c), we have
XL XM = XE + XE .
We would like to generalize this identity to the case where the space Ext1 (L, M)
is of higher dimension. For three objects L, M and N of CQ , let Ext 1 (L, M)N
be the subset of Ext1 (L, M) formed by those morphisms ε : L → M such
that in the triangle
ε
M → E → L → M,
the object E is isomorphic to N (we do not fix an isomorphism). Notice that

this subset is a cone (i.e. stable under multiplication by non zero scalars) in the
vector space Ext1 (L, M).
Proposition 5.13 ([30]). The subset Ext1 (L, M)N is constructible in

Ext1 (L, M). In particular, it is a union of algebraic subvarieties. It is empty for
all but finitely isomorphism classes of objects N .
If k is the field of complex numbers, we denote by χ the Euler characteristic

with respect to singular cohomology with coefficients in a field. If k is an
arbitrary algebraically closed field, we denote by χ the Euler characteristic
with respect to étale cohomology with proper support.
Theorem 5.14 ([30]). Suppose that L and M are objects of CQ such that
Ext1 (L, M) = 0. Then we have
χ (P Ext1 (L, M)N ) + χ (P Ext1 (M, L)N )
XL XM = XN ,
N
χ (P Ext1 (L, M))
where the sum is taken over all isomorphism classes of objects N of CQ .

122 Bernhard Keller
Notice that in the theorem, the objects L and M may be decomposable

so that XL and XM will not be cluster variables in general and the XN do
not form a linearly independent set in the cluster algebra. Thus, the formula
should be considered as a relation rather than as an alternative definition for the
multiplication of the cluster algebra. Notice that it nevertheless bears a close
resemblance to the product formula in a dual Hall algebra: For two objects L
and M in a finitary abelian category of finite global dimension, we have
| Ext1 (L, M)N |
[L] ∗ [M] = [N ] ,
[N]
| Ext1 (L, M)|
where the brackets denote isomorphism classes and the vertical bars the cardi-
nalities of the underlying sets, cf. Proposition 1.5 of [119].
6. Categorification via cluster categories: the acyclic case

6.1. Categorification
Let Q be a connected finite quiver without oriented cycles with vertex set
{1, . . . , n}. Let k be an algebraically closed field. We have seen in section 5.4
how to define the bounded derived category DQ . We still have a fully faithful
functor from the mesh category of ZQ to the category of indecomposables
of DQ but this functor is very far from being essentially surjective. In fact,
its image does not even contain the injective indecomposable kQ-modules.
The methods of the preceding section therefore do not generalize but most of
the results continue to hold. The derived category DQ still has a Serre functor
(the total left derived functor of the tensor product functor ? ⊗B D(kQ)). We
can form the cluster category
CQ = DQ /(S −1 2 )Z
as before and it is still a triangulated category in a canonical way such that
the projection π : DQ → CQ becomes a triangle functor [90]. Moreover, the
decomposition theorem 5.1 holds for CQ and each object L of CQ decomposes
into a direct sum
n
L = π(M) ⊕ π(Pi )mi
i=1
for some module M and certain multiplicities mi , 1 ≤ i ≤ n, cf. [10]. We put

n
XL = CC(M) ximi ,
i=1
where CC(M) is defined as in section 5.3 Notice that in general, XL can only
be expected to be an element of the fraction field Q(x1 , . . . , xn ), not of the
cluster algebra AQ inside this field. (The exponents in the formula for XL are
perhaps more transparent in equation 7.5.1 below).
Theorem 6.1. Let Q be a finite quiver without oriented cycles with vertex set
{1, . . . , n}.
a) The map L
→ XL induces a bijection from the set of isomorphism classes
of rigid indecomposables of the cluster category CQ onto the set of cluster
variables of the cluster algebra AQ .
b) Under this bijection, the clusters correspond exactly to the cluster-tilting
sets, i.e. the sets T1 , . . . , Tn of rigid indecomposables such that
Ext 1 (Ti , Tj ) = 0
for all i, j .
c) For a cluster-tilting set T1 , . . . , Tn , the quiver of the endomorphism algebra
of the sum of the Ti does not have loops nor 2-cycles and encodes the
exchange matrix of the [68] seed containing the corresponding cluster.
d) If L and M are rigid indecomposables such that the space Ext1 (L, M) is
one-dimensional, then we have the generalized exchange relation
XL XM = XB + XB (6.1.1)
where B and B are the middle terms of ‘the’ non split triangles
L /B /M / L and M / B /L / M .
Parts a), b) and d) of the theorem are proved in [29] and part c) in
[11]. The proofs build on work by many authors notably Buan-Marsh-
Reiten-Todorov [12] Buan-Marsh-Reiten [11], Buan-Marsh-Reineke-Reiten-
Todorov [10], Marsh-Reineke-Zelevinsky [106], . . . and especially on Caldero-
Chapoton’s explicit formula for XL proved in [27] for orientations of simply
laced Dynkin diagrams. Another crucial ingredient of the proof is the Calabi-
Yau property of the cluster category. An alternative proof of part c) was given
by A. Hubery [79] for quivers whose underlying graph is an extended simply
laced Dynkin diagram.
We describe the main steps of the proof of a). The mutation of cluster-tilting
sets is defined using the construction of section 5.10.
1) If T is a cluster-tilting object, then the quiver QT of its endomorphism

algebra does not have loops or 2-cycles. If T is obtained from T by mutation
124 Bernhard Keller
at the summand T1 , then the quiver QT of the endomorphism algebra of T

is the mutation at the vertex 1 of the quiver QT , cf. [11].
2) Each rigid indecomposable is contained in a cluster-tilting set. Any two
cluster-tilting sets are linked by a finite sequence of mutations. This is
deduced in [10] from the work of Happel-Unger [76].
3) If (T1 , T1∗ ) is an exchange pair and
T1∗ → E → T1 → T1∗ and T1 → E → T1∗ → T1
are the exchange triangles, then we have
XT1 XT1∗ = XE + XE .
This is shown in [29].
It follows from 1)-3) that the map L → XL does take rigid indecomposables to
cluster variables and that each cluster variable is obtained in this way. It remains
to be shown that a rigid indecomposable L is determined up to isomorphism
by XL . This follows from
4) If M is a rigid indecomposable module, the denominator of XM is x1d1 . . . xndn ,

cf. [29].
Indeed, a rigid indecomposable module M is determined, up to isomorphism,

by its dimension vector.
We sum up the relations between the cluster algebra and the cluster category
in the following table
cluster algebra cluster category

multiplication direct sum
addition ?
cluster variables rigid indecomposables
clusters cluster-tilting sets
mutation mutation
exchange relation exchange triangles
xx ∗ = m + m Tk → M → Tk∗ → Tk
Tk∗ → M → Tk → Tk∗
6.2. Two applications

Theorem 6.1 does shed new light on cluster algebras. In particular, thanks to
the theorem, Caldero and Reineke [31] have made significant progress towards
the
Conjecture 6.2. Suppose that Q does not have oriented cycles. Then all cluster
variables of AQ belong to N[x1± , . . . , xn± ].
This conjecture is a consequence of a general conjecture of Fomin-

Zelevinsky [50], which here is specialized to the case of cluster algebras
associated with acyclic quivers, for cluster expansions in the initial cluster.
Caldero-Reineke’s work in [31] is based on Lusztig’s [105] and in this sense it
does not quite live up to the hopes that cluster theory ought to explain Lusztig’s
results. Notice that in [31], the above conjecture is stated as a theorem. How-
ever, a gap in the proof was found by Nakajima [109]: the authors incorrectly
identify their parameter q with Lusztig’s parameter v, whereas the correct
√
identification is v = − q.
Here are two applications to the exchange graph of the cluster algebra
associated with an acyclic quiver Q:
Corollary 6.3 ([29]).

a) For any cluster variable x, the set of seeds whose clusters contain x form a
connected subgraph of the exchange graph.
b) The set of seeds whose quiver does not have oriented cycles form a connected
subgraph (possibly empty) of the exchange graph.
For acyclic cluster algebras, parts a) and b) confirm conjecture 4.14 parts
(3) and (4) by Fomin-Zelevinsky in [53]. By b), the cluster algebra associated
with a quiver without oriented cycles has a well-defined cluster-type.
6.3. Cluster categories and singularities

The construction of cluster categories may seem a bit artificial. Nevertheless,
cluster categories do occur ‘in nature’. In particular, certain triangulated cat-
egories associated with singularities are equivalent to cluster categories. We
illustrate this on the following example: Let the cyclic group G of order 3 act
on a three-dimensional complex vector space V by scalar multiplication with a
primitive third root of unity. Let S be the completion at the origin of the coordi-
nate algebra of V and let R = S G the fixed point algebra, corresponding to the
completion of the singularity at the origin of the quotient V //G. The algebra R
is a Gorenstein ring, cf. e.g. [128], and an isolated singularity of dimension 3, cf.
e.g. Corollary 8.2 of [81]. The category CM(R) of maximal Cohen-Macaulay
modules is an exact Frobenius category and its stable category CM(R) is a
triangulated category. By Auslander’s results [6], cf. Lemma 3.10 of [130], it
is 2-Calabi Yau. One can show that it is equivalent to the cluster category CQ
126 Bernhard Keller
for the quiver

//
Q: 1 /2
by an equivalence which takes the cluster-tilting object T = kQ to S considered
as an R-module. This example can be found in [91], where it is deduced from
an abstract characterization of cluster categories. A number of similar examples
can be found in [25] and [93].
7. Categorification via 2-Calabi-Yau categories

The extension of the results of the preceding sections to quivers containing
oriented cycles is the subject of ongoing research, cf. for example [39] [61] [8].
Here we present an approach based on the fact that many arguments developed
for cluster categories apply more generally to suitable triangulated categories,
whose most important property it is to be Calabi-Yau of dimension 2. As an
application, we will sketch a proof of the periodicity conjecture in section 8
(details will appear elsewhere [86]).
7.1. Definition and main examples

Let k be an algebraically closed field and let C a triangulated category with sus-
pension functor where all idempotents split (i.e. each idempotent endomor-
phism e of an object M is the projection onto M1 along M2 in a decomposition
M = M1 ⊕ M2 ). We assume that
1) C is Hom-finite (i.e. we have dim C(L, M) < ∞ for all L, M in C) and the
decomposition theorem 5.1 holds for C;
2) C is 2-Calabi-Yau, i.e. we are given bifunctorial isomorphisms
∼
DC(L, M) → C(M, 2 L) , L, M ∈ C;
3) C admits a cluster-tilting object T , i.e.
a) T is the sum of pairwise non-isomorphic indecomposables,
b) T is rigid and
c) for each object L of C, if Ext1 (T , L) vanishes, then L belongs to the
category add(T ) of direct factors of finite direct sums of copies of T .
If all these assumptions hold, we say that (C, T ) is a 2-Calabi-Yau category
with cluster tilting object. If Q is a finite quiver, we say that a 2-Calabi-Yau
category C with cluster-tilting object T is a 2-Calabi-Yau realization of Q if Q
is the quiver of the endomorphism algebra of T .
For example, if C is the cluster category of a finite quiver Q without oriented

cycles, conditions 1) and 2) hold and an object T is cluster-tilting in the above
sense iff it is the direct sum of a cluster-tilting set T1 , . . . , Tn , where n is the
number of vertices of Q, cf. [10]. The ‘initial’ cluster-tilting object in this case is
T = kQ (the image in CQ of the free module of rank one) and (C, kQ) is the
canonical 2-Calabi-Yau realization of the quiver Q (without oriented cycles).
The second main class of examples comes from the work of Geiss-Leclerc-
Schröer: Let be a simply laced Dynkin diagram, a quiver with underlying
graph and the doubled quiver obtained from by adjoining an arrow
α ∗ : j → i for each arrow α : i → j . The preprojective algebra = () is
the quotient of the path algebra of by the ideal generated by the relator

αα ∗ − α ∗ α.
α∈Q1
is the quiver
For example, if
1
α /2 β
/3,
then is the quiver

α / β
/
1o ∗
2o 3
α β∗
and the ideal generated by the above sum of commutators is also generated by
the elements
α ∗ α, αα ∗ − β ∗ β, ββ ∗ .
It is classical, cf. e.g. [115], that the algebra = () is a finite-dimensional
(!) selfinjective algebra (i.e. is injective as a right -module over itself). Let
mod denote the category of k-finite-dimensional right -modules. The stable
module category mod is the quotient of mod by the ideal of all morphisms
factoring through a projective module. This category carries a canonical trian-
gulated structure (like any stable module category of a self-injective algebra):
The suspension is constructed by choosing exact sequences of modules
0 → L → I L → L → 0
where I L is injective (but not necessarily functorial in L; the object L
becomes functorial in L when we pass to the stable category). The triangles are
by definition isomorphic to standard triangles obtained from exact sequences
of modules as follows: Let
i p
0→L→M→N →0
128 Bernhard Keller
be a short exact sequence of mod . Choose a commutative diagram
0 /L i /M p
/N /0
1 e

0 /L / IL / L /0
Then the image of (i, p, e) is a standard triangle in the stable module category.
As shown in [36], the stable module category C = mod is 2-Calabi-Yau and
it is easy to check that assumption 1) holds.
Theorem 7.1 (Geiss-Leclerc-Schröer). The category C = mod admits a
cluster-tilting object T such that the quiver of End(T ) is obtained from that of
the category of indecomposable k-modules by deleting the injective vertices
and adding an arrow v → τ v for each non projective vertex v.
Here, by the quiver of the category of indecomposable k-modules, we
mean the full subquiver of the repetition ZQ which is formed by the vertices
corresponding to modules (complexes concentrated in degree 0). Thus for
= A5 , this quiver is as follows:
P5? = ?I1
???

? P4 ?? ? I2 ???
?? ?
P3 ?7? ? I3 ???
? ??? ??? ?
?
P ? ?
6 10
? ? ? I4 ???
? ???
2
? ??? ??? ?

P1 5 9 12 I5
where we have marked the indecomposable projectives Pi and the indecom-
posable injectives Ij . If we remove the vertices corresponding to the indecom-
posable injectives (and all the arrows incident with them) and add an arrow
v → τ v for each vertex not corresponding to an indecomposable projective,
we obtain the following quiver
?0?
???

o
? 1 ?? ?2?
?? ???

o o
? 3 ?? ? 4 ?? ? ???
5
?? ?? ?

6o 7o 8o 9
In a series of papers [64] [66] [65] [62] [61] [67] [63], Geiss-Leclerc-Schröer
have obtained remarkable results for a class of quivers which are important
in the study of (dual semi-)canonical bases. They use an analogue [67] of
the Caldero-Chapoton map due ultimately to Lusztig [104]. The class they
consider has been further enlarged by Buan-Iyama-Reiten-Scott [7]. Thanks
to their results, an analogue of Caldero-Chapoton’s formula and a weakened
version of theorem 6.1 was proved in [57] for an even larger class.
7.2. Calabi-Yau reduction

Suppose that (C, T ) is a 2-Calabi-Yau realization of a quiver Q so that for
1 ≤ i ≤ n, the vertex i of Q corresponds to the indecomposable summand Ti
of T . Let J be a subset of the set of vertices of Q and let Q be the quiver
obtained from Q by deleting all vertices in J and all arrows incident with one
of these vertices. Let U be the full subcategory of C formed by the objects U
such that Ext1 (Tj , U ) = 0 for all j ∈ J . Note that all Ti , 1 ≤ i ≤ n, belong
to U. Let Tj |j ∈ J denote the ideal of U generated by the identities of the
objects Tj , j ∈ J . By imitating the construction of the triangulated structure
on a stable category, one can endow the quotient
C = U/Tj |j ∈ J
with a canonical structure of triangulated category, cf. [81].
Theorem 7.2 (Iyama-Yoshino [81]). The pair (C , T ) is a 2-Calabi-Yau real-

ization of the quiver Q . Moreover, the projection U → C induces a bijection
between the cluster-tilting sets of C containing the Tj , j ∈ J , and the cluster-
tilting sets of C .
7.3. Mutation
Let (C, T ) be a 2-Calabi-Yau category with cluster-tilting object. Let T1 be an
indecomposable direct factor of T .
Theorem 7.3 (Iyama-Yoshino [81]). Up to isomorphism, there is a unique

indecomposable object T1∗ not isomorphic to T1 such that the object μ1 (T )
obtained from T by replacing the indecomposable summand T1 with T1∗ is
cluster-tilting.
We call μ1 (T ) the mutation of T at T1 . If C is the cluster category of

a finite quiver without oriented cycles, this operation specializes of course
to the one defined in section 5.10. However, in general, the quiver Q of the
130 Bernhard Keller
endomorphism algebra of T may contain loops and 2-cycles and then the quiver
of the endomorphism algebra of μ1 (T ) is not determined by Q. Let us illustrate
this phenomenon on the following example (taken from Proposition 2.6 of
[25]): Let C be the orbit category of the bounded derived category DD 6 under
the action of the autoequivalence τ 2 . Then C satisfies the assumptions 1) and
2) of section 7.1. Its category of indecomposables is equivalent to the mesh
category of the quiver
C? C ?C?
?? ? ???? ???
?
/ 5 / B / 11 / B / 17
B ? ?? ? ?? ?
?? ??

1?
? ???
7 13
?? ? ???
? ? ?
?4? ? 10 ? ? 16
??? ???

A A A
where the vertices labeled A, 1, B, C, 4, 5 on the left have to be identified
with the vertices labeled A, 13, B, C, 16, 17 on the right. In this case, there
are exactly 6 indecomposable rigid objects, namely A, B, C, A , B and C .
There are exactly 6 cluster-tilting sets. The following is the exchange graph: Its
vertices are cluster-tilting sets (we write AC instead of {A, C}) and its edges
represent mutations.
AC AB6
66
66
6
B C6 CB
66
66
6
B A C A
The quivers of the endomorphism algebras are as follows:
◦o /◦
AC, AB : d
B C, C B : ◦o /◦
/◦
B A ,C A : :◦o
In the setting of the above theorem, there are still exchange triangles as in
theorem 5.12 but their description is different: Let T be the full subcategory
of C formed by the direct sums of indecomposables Ti , where i is different
from k. A left T -approximation of Tk is a morphism f : Tk → T with T
in T such that any morphism from Tk to an object of T factors through f .

A left T -approximation f is minimal if for each endomorphism g of T , the
equality gf = f implies that g is invertible. Dually, one defines (minimal) right
T -approximations. It is not hard to show that they always exist.
Theorem 7.4 (Iyama-Yoshino [81]). If (T1 , T1∗ ) is an exchange pair, there are
non split triangles, unique up to isomorphism,
f g
T1 → E → T1∗ → T1 and T1∗ → E → T1 → T1∗
such that f is a minimal left T -approximation and g a minimal right T -

approximation.
7.4. Simple mutations, reachable cluster-tilting objects

Let (C, T ) be a 2-CY category with cluster-tilting object and T1 an indecom-
posable direct summand of T . Let (T1 , T1∗ ) be the corresponding exchange pair
and
T1∗ → E → T1 → T1∗
the exchange triangle. The long exact sequence induced in C(T1 , ?) by this
triangle yields a short exact sequence
CT (T1 , T1 ) → C(T1 , T1 ) → Ext1 (T1 , T1∗ ) → 0 ,
where the leftmost term is the space of those endomorphisms of T1 which factor
through a sum of copies of T /T1 . Now the algebra C(T1 , T1 ) is local and its
residue field is k (since k is algebraically closed). We deduce the following
lemma.
Lemma 7.5. The quiver of the endomorphism algebra of T does not have
a loop at the vertex corresponding to T1 iff we have dim Ext1 (T1 , T1∗ ) = 1 iff
Ext1 (T1 , T1∗ ) is a simple module over C(T1 , T1 ). In this case, in the exchange
triangles
T1∗ → E → T1 → T1∗ and T1 → E → T1∗ → T1 ,
we have

E= Ti and E = Tj .
arrows arrows
i→1 1→j
We say that the mutation at T1 is simple if the conditions of the lemma

hold. If C is the cluster category of a finite quiver without oriented cycles, all
132 Bernhard Keller
mutations in C are simple, by part c) of theorem 6.1. By a theorem of Geiss-

Leclerc-Schröer, if C is the stable module category of a preprojective algebra
of Dynkin type, the quiver of any cluster-tilting object in C does not have loops
nor 2-cycles. So again, all mutations are simple in this case.
Theorem 7.6 (Buan-Iyama-Reiten-Scott [7]). Suppose that the quivers Q and

Q of the endomorphism algebras of T and T = μ1 (T ) do not have loops nor
2-cycles. Then Q is the mutation of Q at the vertex 1.
We define a cluster-tilting object T to be reachable from T if there is a

sequence of mutations
T = T (0) T (1) . . . T (N) = T
such that the quiver of End(T (i) ) does not have loops nor 2-cycles for all
1 ≤ i ≤ N . We define a rigid indecomposable of C to be reachable from T if it
is a direct summand of a reachable cluster-tilting object.
Corollary 7.7. If a cluster-tilting object T is reachable from T , then the quiver

of the endomorphism algebra of T is mutation-equivalent to the quiver of the
endomorphism algebra of T . If C is a cluster-category or the stable module
category of the preprojective algebra of a Dynkin diagram, then all quivers
mutation-equivalent to Q are obtained in this way.
7.5. Combinatorial invariants

Let (C, T ) be a 2-Calabi-Yau category with cluster-tilting object. Let T be the
full subcategory whose objects are all direct factors of finite direct sums of
copies of T . Notice that T is equivalent to the category of finitely generated
projective modules over the endomorphism algebra of T . Let K0 (T ) be the
Grothendieck group of the additive category T . Thus, the group K0 (T ) is
free abelian on the isomorphism classes of the indecomposable summands
of T .
Lemma 7.8 (Keller-Reiten [92]). For each object L of C, there is a triangle
T1 → T0 → L → T1
such that T0 and T1 belong to T . The difference
[T0 ] − [T1 ]
considered as an element of K0 (T ) does not depend on the choice of this

triangle.
In the situation of the lemma, we define the index ind(L) of L as the element
[T0 ] − [T1 ] of K0 (T ).
Theorem 7.9 (Dehy-Keller [37]).

a) Two rigid objects are isomorphic iff their indices are equal.
b) The indices of the indecomposable summands of a cluster-tilting object
form a basis of K0 (T ). In particular, all cluster-tilting objects have the
same number of pairwise non isomorphic indecomposable summands.
Let B be the endomorphism algebra of T . For two finite-dimensional right

B-modules L and M put
L, Ma = dim Hom(L, M) − dim Ext1 (L, M)

− dim Hom(M, L) + dim Ext1 (M, L).
This is the antisymmetrization of a truncated Euler form. A priori it is defined

on the split Grothendieck group of the category mod B (i.e. the quotient of
the free abelian group on the isomorphism classes divided by the subgroup
generated by all relations obtained from direct sums in mod B).
Proposition 7.10 (Palu [112]). The form , a descends to an antisymmetric

form on K0 (mod B). Its matrix in the basis of the simples is the antisymmetric
matrix associated with the quiver of B (loops and 2-cycles do not contribute to
this matrix).
Let T1 , . . . , Tn be the pairwise non isomorphic indecomposable direct sum-

mands of T . For L ∈ C, we define the integer gi (L) to be the multiplicity of
[Ti ] in the index ind(L), 1 ≤ i ≤ n, and we define the element XL of the field
Q(x1 , . . . , xn ) by

n
n
Si ,ea
XL =
g (L)
xi i χ (Gre (Ext1 (T , L))) xi , (7.5.1)
i=1 e i=1
where Si is the simple quotient of the indecomposable projective B-module

Pi = Hom(T , Ti ). Notice that we have XT i = xi , 1 ≤ i ≤ n. If C is the cluster-
category of a finite quiver Q without oriented cycles and T = kQ, then we have
XL = XL in the notations of section 6 and the formula for XL is essentially
another expression for the Caldero-Chapoton formula.
Now let Q be the quiver of the endomorphism algebra of T in C and let AQ
be the associated cluster algebra.
134 Bernhard Keller
Theorem 7.11 (Palu [112]). If L and M are objects of C such that Ext1 (L, M)
is one-dimensional, then we have
XL XM

= XE + XE ,
where
L → E → M → L and M → E → L → M
are ‘the’ two non split triangles. Thus, if L is a rigid indecomposable reachable
from T , then XL is a cluster variable of AQ .
Corollary 7.12. Suppose that C is the cluster category CQ of a finite quiver

Q without oriented cycles. Let AQ ⊂ Q(x1 , . . . , xn ) be the associated cluster
algebra and x ∈ AQ a cluster variable. Let L ∈ CQ be the unique (up to
isomorphism) indecomposable rigid object such that x = XL . Let u1 , . . . , un
be an arbitrary cluster of AQ and T1 , . . . , Tn the cluster-tilting set such that
ui = XTi , 1 ≤ i ≤ n. Then the expression of x as a Laurent polynomial in
u1 , . . . , un is given by
x = XL (u1 , . . . , un ).
The expression for XL makes it natural to define the polynomial FL ∈
Z[y1 , . . . yn ] by

n
FL =
e
χ (Gre (Ext1 (T , L))) yj j . (7.5.2)
e j =1
We then have
! n "

n
n
XL xi i FL
g (L)
= xibi1 , . . . , xibin .
i=1 i=1 i=1
The polynomial FL is related to Fomin-Zelevinsky’s F -polynomials [54] [55],

as we will see below.
7.6. More mutants categorified

Let Q be a finite quiver without loops nor 2-cycles with vertex set {1, . . . , n}.
Let Tn be the regular n-ary tree: Its edges are labeled by the integers 1, . . . ,
n such that the n edges emanating from each vertex carry different labels. Let
t0 be a vertex of Tn . To each vertex t of Tn we associate a seed (Qt , xt ) (cf.
section 3.2) such that at t = t0 , we have Qt = Q and xt = {x1 , . . . , xn } and
whenever t is linked to t by an edge labeled i, we have (Qt , xt ) = μi (Qt , xt ).
Now assume that Q admits a 2-Calabi-Yau realization (C, T ). According to

the mutation theorem 7.6, we can associate a cluster-tilting object Tt to each
vertex t of Tn such that at t = t0 , we have Tt = T and that whenever t is linked
to t by an edge labeled i, we have Tt = μi (Tt ).
Now let
i1 i2 iN
t0 t1 ... tN
be a path in Tn and suppose that for each 1 ≤ i ≤ N , the quiver of the
endomorphism algebra of Tti does not have loops nor 2-cycles. Then it follows
by induction from theorem 7.6 that the quiver of the endomorphism algebra
of Tti is Qti , 1 ≤ i ≤ N , and from theorem 7.11 that the cluster xti equals the
image under L
→ XL of the set of indecomposable direct factors of Tti .
Following [55], let us consider three other pieces of data associated with
each vertex t of Tn :
r the tropical Y -variables y1,t , . . . , yn,t ,
r the F -polynomials F1,t , . . . , Fn,t ,
r the (non tropical) Y -variables Y1,t , . . . , Yn,t .
Here the tropical Y -variables are monomials in the indeterminates y1 , . . . , yn
and their inverses. At t = t0 , we have yi,t = yi , 1 ≤ i ≤ n. If t is linked to t
by an edge labelled i, then

−1
yi,t if j = i
yj,t = [bij ]+ −bij
yj,t yi,t (yi,t ⊕ 1) if j = i.
Here, (bij ) is the antisymmetric matrix associated with Qt , for an integer a,
we write [a]+ for max(a, 0), and, for a monomial m which is the product of
e e
powers yj j , 1 ≤ j ≤ n, we write m ⊕ 1 for the product of the factors yj j with
negative exponents ej . Notice that [bij ]+ is the number of arrows from i to j
in Qt .
The F -polynomials lie in Z[y1 , . . . , yn ]. At t = t0 , they all equal 1. If t is
linked to t by an edge labeled i, then
Fj,t = Fj,t if j = i ,
⎛ ⎞
1 ⎝ cli [bij ]+ −cli [−bij ]+ ⎠
n n
Fi,t = y F + yl Fj,t ,
Fi,t c >0 l j =1 j,t c <0 j =1
li li

where the cli are the exponents in the tropical Y -variables yi,t = nl=1 ylcli and
B = (bij ) is the antisymmetric matrix associated with Qt .
Finally, the non tropical Y -variables Yj,t lie in the field Q(y1 , . . . , yn ). At
t = t0 , we have Yj,t = yj , 1 ≤ j ≤ n, and if t and t are linked by an edge
136 Bernhard Keller
labeled i, then

−1
Yi,t if j = i
Yj,t = [bij ]+ −bij
(7.6.1)
Yj,t Yi,t (Yi,t + 1) if j = i.
The principal extension Q of Q is the quiver obtained from Q by adding,

for each vertex i of Q, a new vertex i and a new arrow i → i. We assume
T) of Q.
that we are given a 2-Calabi-Yau realization (C, Let U ⊂ Cbe the full
subcategory of the objects U such that Ext1 (Ti , U ) = 0 for each new vertex i .
Then, according to the theorem on Calabi-Yau reductions (cf. section 7.2), the
quotient category
U/Ti |i ∈ Q0
together with the image of T is a 2-Calabi-Yau realization of the quiver Q. We

assume that (C, T ) is this realization. Now let
i1 i2 iN
t0 t1 ... tN
be a path in Tn and suppose that for each 1 ≤ i ≤ N , the quiver of the endo-
morphism algebra of Tti does not have loops nor 2-cycles. Let t = tN and let
T1 , . . . , Tn be the indecomposable summands of T = Tt . Recall from part b)
of theorem 7.9 that the indices ind(Tl ), 1 ≤ l ≤ n, form a basis of the group
K0 (T ), where T is the full subcategory of C formed by the direct summands
of finite direct sums of copies of T .
Theorem 7.13.
a) The exchange matrix Bt = (bij ) associated with t is the antisymmetric
matrix associated with the quiver of the endomorphism algebra of Tt .
c
b) We have yj,t = nl=1 yl lj , 1 ≤ j ≤ n, where clj is defined by

n
[Tj ] = clj ind(Tl ).
l=1
c) We have Fj,t = FT , 1 ≤ j ≤ n, where FT is defined by equation 7.5.2.

j j
d) We have

n
b
Yj,t = yj,t Fi,tij .
i=1
Here part a) follows by induction from theorem 7.6, part b) is proved by

applying theorem 7.6 and Theorem 12 of [111] to the category C, part c) is
Theorem 5.3 of [57] and part d) is Proposition 3.12 of [55].
7.7. 2-CY categories from algebras of global dimension 2

The results of the preceding sections only become interesting if we are able to
construct 2-Calabi-Yau realizations for large classes of quivers. In the case of a
quiver without oriented cycles, this problem is solved by the cluster category.
The results of Geiss-Leclerc-Schröer provide another large class of quivers
admitting 2-Calabi-Yau realizations. Here, we will exhibit a construction which
generalizes both cluster categories of quivers without oriented cycles and the
stable categories of preprojective algebras of Dynkin diagrams.
Let k be an algebraically closed field and A a finite-dimensional k-algebra
of global dimension ≤ 2. For example, A can be the path algebra of a finite
quiver without oriented cycles. Let DA be the bounded derived category of
the category mod A of k-finite-dimensional right A-modules. It admits a Serre
functor, namely the total derived functor of the tensor product ? ⊗A DA with
the k-dual bimodule of A considered as a bimodule over itself. We can form
the orbit category
DA /(S −1 2 )Z .
This is a k-linear category endowed with a suspension functor (induced by

) but in general it is no longer triangulated. Nevertheless, one can show that
it embeds fully faithfully into a ‘smallest triangulated overcategory’ [90]. We
denote this overcategory by CA and call it the generalized cluster category
of A.
Theorem 7.14 (Amiot [1]). If the functor
2 (?, DA) : mod A → mod A

TorA
is nilpotent (i.e. some power of it vanishes), then CA is Hom-finite and 2-Calabi-

Yau. Moreover, the image T of A in CA is a cluster-tilting object. The quiver of
its endomorphism algebra is obtained from that of A by adding, for each pair
of vertices (i, j ), a number of arrows equal to
op
dim TorA
2 (Sj , Si )
op
from i to j , where Sj is the simple right module associated with j and Si the
simple left module associated with i.
We consider two classes of examples obtained from this theorem: First, let
be a simply laced Dynkin diagram and k the path algebra of a quiver with
i.e. the endomorphism
underlying graph . Let A the Auslander algebra of k ,
algebra of the direct sum of a system of representatives of the indecompos-
able B-modules modulo isomorphism. Then it is not hard to check that the
138 Bernhard Keller
•o ◦ /•o
@◦
@ ^===
===
• 5o / o
55 {= ◦ 5iR5R RR =•==55 {= ◦ 55
5 {{{{ 55 RRR =5=5 {{{
RRR {
55

o
◦O NN ◦O / o
◦O NN ◦
NNN
NNN ppppp
pp NNNN O
& xp N N&
•o p 8 ◦ fNNN /•o
p8 ◦
pp p NNN pp p
pp p p NN pp pp
◦o ◦ /◦o ◦
Figure 4. The quiver A4 ⊗ D

5
assumptions of the theorem hold. The quiver of A is simply the quiver of

the category of indecomposables of k and the minimal relations correspond
to its meshes. For example, if we choose = A4 with the linear orienta-
tion, the quiver of the endomorphism algebra of the image T of A in CA is
the quiver obtained from Geiss-Leclerc-Schröer’s construction at the end of
section 7.1.
As a second class of examples, we consider an algebra A which is the tensor
product kQ ⊗k kQ of two path algebras of quivers Q and Q without oriented
cycles. Such an algebra is clearly of global dimension ≤ 2. Let us assume that Q
and Q are moreover Dynkin quivers. Then it is not hard to check that the functor
Tor2 (?, DA) is indeed nilpotent. Thus, the theorem applies. Another elementary
op
exercise in homological algebra shows that the space TorA 2 (Sj,j , Si,i ) is at most
one-dimensional and that it is non zero if and only if there is an arrow i → j in
Q and an arrow i → j in Q . This immediately yields the shape of the quiver
of the endomorphism algebra of the image T of A in CA : It is the tensor product
Q ⊗ Q obtained from the product Q × Q by adding an arrow (j, j ) → (i, i )
for each pair of arrows i → j of Q and i → j of Q . For example, for
suitable orientations of A4 and D5 , we obtain the quiver of figure 4. Notice
that if we perform mutations at the six vertices of the form (i, i ), where i
is a sink of Q = A4 and i a source of Q = D 5 (they are marked by •), we
obtain the quiver of figure 5 related to the periodicity conjecture for (A4 , D5 ),
cf. below.
8. Application: The periodicity conjecture

Let and be two Dynkin diagrams with vertex sets I and I . Let A and A
be the incidence matrices of and , i.e. if C is the Cartan matrix of and J
◦J /◦o ◦J /◦

◦ X1 / ◦ 1o
◦ X1 / ◦ 1
11 11 11 11
1 1
1 1

◦o ◦O /◦o ◦O

◦O /◦o ◦O /◦

◦o ◦ /◦o ◦
5
Figure 5. The quiver A4 D
the identity matrix of the same format, then A = 2J − C. Let h and h be the
Coxeter numbers of and .
The Y -system of algebraic equations associated with the pair of Dynkin
diagrams (, ) is a system of countably many recurrence relations in the
variables Yi,i ,t , where (i, i ) is a vertex of × and t an integer. The system
reads as follows:

j ∈I (1 + Yj,i ,t )
aij
Yi,i ,t−1 Yi,i ,t+1 = . (8.0.1)
−1 ai j
j ∈I (1 + Yi,j ,t )
Periodicity Conjecture 8.1. All solutions to this system are periodic in t of

period dividing 2(h + h ).
Here is an algebraic reformulation: Let K be the fraction field of the ring
of integer polynomials in the variables Yii , where i runs through the set of
vertices I of and i through the set of vertices I of . Since is a tree,
the set I is the disjoint union of two subsets I+ and I− such that there are no
edges between any two vertices of I+ and no edges between any two vertices
of I− . Analogously, I is the disjoint union of two sets of vertices I+ and I− .
For a vertex (i, i ) of the product I × I , define ε(i, i ) to be 1 if (i, i ) lies in
I+ × I+ ∪ I− × I− and −1 otherwise. For ε = ±1, define an automorphism τε
of K by τε (Yi,i ) = Yi,i if ε(i, i ) = ε and

τε (Yii ) = Yii−1
(1 + Yj i )aij (1 + Yij−1 )−ai j (8.0.2)
j j
if ε(i, i ) = ε. Finally, define an automorphism ϕ of K by

ϕ = τ+ τ− . (8.0.3)
Then, as in [54], it is not hard to check that the periodicity conjecture holds iff
the order of the automorphism ϕ is finite and divides h + h .
140 Bernhard Keller
The conjecture was formulated

r by Al. B. Zamolodchikov for (, A1 ), where is simply laced [131, (12)];
r by Ravanini-Valleriani-Tateo for (, ), where and are simply laced
or ‘tadpoles’ [113, (6.2)];
r by Kuniba-Nakanishi for (, An ), where is not necessarily simply laced
[100, (2a)], see also Kuniba-Nakanishi-Suzuki [101, B.6]; notice however
that in the non simply laced case, the Y -system given by Kuniba-Nakanishi
is different from the one above and that they conjecture periodicity with
period dividing twice the sum of the dual Coxeter numbers;
It was proved
r for (An , A1 ) by Frenkel-Szenes [56] (who produced explicit solutions) and
by Gliozzi-Tateo [72] (via volumes of threefolds computed using triangula-
tions),
r by Fomin-Zelevinsky [54] for (, A1 ), where is not necessarily simply
laced (via the cluster approach and a computer check for the exceptional
types; a uniform proof can be given using [129]),
r for (An , Am ) by Volkov [127], who exhibited explicit solutions using cross
ratios, and by Szenes [124], who interpreted the system as a system of flat
connections on a graph; an equivalent statement was proved by Henriques
[77];
r by Hernandez-Leclerc for (An , A1 ) using representations of quantum affine
algebras (which yield formulas for solutions in terms of q-characters). They
expect to treat (An , ) similarly, cf. [78].
Theorem 8.2 ([86]). The periodicity conjecture 8.1 is true.
Let us sketch a proof based on 2-Calabi-Yau categories (details will appear
elsewhere [86]): First, using the folding technique of Fomin-Zelevinsky’s [54]
one reduces the conjecture to the case where both and are simply laced.
Thus, from now on, we assume that and are simply laced.
First step: We choose quivers Q and Q whose underlying graphs are and
. We assume that Q and Q are alternating, i.e. that each vertex is either a
source or a sink. If i is a vertex of Q or Q , we put ε(i) = 1 if i is a source and
ε(i) = −1 if i is a sink. For example, we can consider the following quivers
A4 : 1 o 2 /3o 4 ,
o4
o
w oooo
5 : 1 o
D 2 / 3 gOO
OOO
O
5.
We define the square product QQ to be the quiver obtained from Q × Q by

reversing all the arrows in the full subquivers of the form {i} × Q and Q × {i },
where i is a sink of Q and i a source of Q . The square product of the above
quivers A4 and D 5 is depicted in figure 5. The initial Y -seed associated with
QQ is the pair y0 formed by the quiver QQ and the family of variables

Yi,i , (i, i ) ∈ I × I . We can apply mutations to it using the quiver mutation rule
and the mutation rule for (non tropical) Y -variables given in equation 7.6.1.
A general construction. Let R be a quiver and v a sequence of vertices
v1 , . . . , vN of R. We assume that the composed mutation
μv = μvN . . . μv2 μv1
transforms R into itself. Then clearly the same holds for the inverse sequence
μ−1
v = μv1 μv2 . . . μvN .
Now the restricted Y -pattern associated with R and μv is the sequence of

Y -seeds obtained from the initial Y -seed y0 associated with R by applying all
integral powers of μv . Thus this pattern is given by a sequence of seeds yt ,
t ∈ Z, such that y0 is the initial Y -seed associated with R and, for all t ∈ Z,
yt+1 is obtained from yt by the sequence of mutations μv .
Second step. For two elements σ , σ of {+, −} define the following com-
posed mutation of QQ :

μσ,σ = μ(i,i ) .
ε(i)=σ,ε(i )=σ
Notice that there are no arrows between any two vertices of the index set so that
the order in the product does not matter. Then it is easy to check that μ+,+ μ−,−
and μ−,+ μ+,− both transform QQ into (QQ )op and vice versa. Thus the
composed sequence of mutations
μ = μ−,− μ+,+ μ−,+ μ+,−
transforms QQ into itself. We define the Y -system y associated with QQ
to be the restricted Y -pattern associated with QQ and μ . As in section 8

of [55], one checks that the identity ϕ h+h = 1 follows if one shows that the
system y is periodic of period dividing h + h .
Third step. One checks easily that we have
μ+− (QQ ) = Q ⊗ Q ,
where the tensor product Q ⊗ Q is defined at the end of section 7.7. For the
above quivers A4 and D 5 , the tensor product is depicted in figure 4. Therefore,
the periodicity of the restricted Y -system associated with QQ and μ is
142 Bernhard Keller
equivalent to that of the restricted Y -system associated with Q ⊗ Q and
μ⊗ = μ+,− μ−,− μ+,+ μ−,+ .
Fourth step. As we have seen in section 7.7, the quiver Q ⊗ Q admits a

2-Calabi-Yau realization given by the cluster category CkQ⊗kQ associated with
the tensor product of the path algebras kQ ⊗ kQ . Let T be the initial cluster
tilting object. By theorem 7.3, we can define its iterated mutations
Tt = μt⊗ (T )
for all t ∈ Z. Now we use the
Proposition 8.3. None of the endomorphism quivers occurring in the

sequence of mutated cluster-tilting objects joining T to Tt contains loops nor
2-cycles.
It is not hard to see that (CkQ⊗kQ , T ) is the Calabi-Yau reduction of a

2-Calabi-Yau realization of the principal extension of QQ . Therefore, it
follows from theorem 7.13 that the quiver of the endomorphism algebra of
Tt is Q ⊗ Q and that the Y -variables in the Y -seed y,t can be expressed in
terms of the triangulated category CkQ⊗kQ and the objects T and Tt . Thus, it
suffices to show that Tt is isomorphic to T whenever t is an integer multiple
of h + h .
Fifth step. Let τ ⊗ 1 denote the auto-equivalence of the bounded derived
category of kQ ⊗ kQ given by the total left derived functor of the tensor
product with the bimodule complex ( −1 D(kQ)) ⊗ kQ , where D is duality
over k. It induces an autoequivalence of CkQ⊗kQ which we still denote by τ ⊗ 1.
Define to be the autoequivalence τ −1 ⊗ 1 of CkQ⊗kQ .
t
Proposition 8.4. For each integer t, the image (T ) is isomorphic to Tt .
The proposition is proved by showing that the indices of the two objects are
equal. This suffices by theorem 7.9. Now we conclude thanks to the following
categorical periodicity result:
h+h
Proposition 8.5. The power is isomorphic to the identity functor.
Let us sketch the proof of this proposition: With the natural abuse of notation,
the Serre functor S of the bounded derived category of kQ ⊗ kQ is given by the
‘tensor product’ S ⊗ S of the Serre functors for kQ and kQ . In the generalized
cluster category, the Serre functor becomes isomorphic to the square of the
suspension functor. So we have
S = S ⊗ S = 2 = ⊗
as autoequivalences of CkQ⊗kQ . Now recall from equation 5.5.4 that τ = −1 S.

Thus we get the isomorphism
τ ⊗τ =1
of autoequivalences of CkQ⊗kQ . So we have
τ −1 ⊗ 1 = 1 ⊗ τ.

Now we compute h+h by using the left hand side for the first h factors and
the right hand side for the last h factors:
h+h
= (τ −1 ⊗ 1)h (1 ⊗ τ )h .
But we know from equation 5.5.3 that τ −h = 2 . So we find
h+h
= ( 2 ⊗ 1)(1 ⊗ −2 ) = 2 −2 = 1
as required.
9. Quiver mutation and derived equivalence

Let Q be a finite quiver and i a source of Q, i.e. no arrows have target i. Then
the mutation Q of Q at i is simply obtained by reversing all the arrows starting
at i. In this case, the categories of representations of Q and Q are related
by the Bernstein-Gelfand-Ponomarev reflection functors [19] and these induce
equivalences in the derived categories [73]. We would like to present a similar
categorical interpretation for mutation at arbitrary vertices. For this, we use
recent work by Derksen-Weyman-Zelevinsky [39] and a construction due to
Ginzburg [71]. We first recall the classical reflection functors in a form which
generalizes well.
9.1. A reminder on reflection functors

We keep the above notations. In the sequel, k is a field and kQ is the path
algebra of Q over k. For each vertex j of Q, we write ej for the lazy path
at j , the idempotent of kQ associated with j . We write Pj = ej kQ for the
corresponding indecomposable right kQ-module, Mod kQ for the category of
(all) right kQ-modules and D(kQ) for its derived category.
144 Bernhard Keller
The module categories over kQ and kQ are linked by a pair of adjoint
functors [19]

Mod kQ
O
F0 G0

Mod kQ.
The right adjoint G0 takes a representation V of Qop to the representation V

with Vj = Vj for j = i and where Vi is the kernel of the map

Vj −→ Vi
arrows of Q
i→j
whose components are the images under V of the arrows i → j .

Then the left derived functor of F0 and the right derived functor of G0 are
quasi-inverse equivalences [73]

DkQ
O
F G

DkQ.
The functor F sends the indecomposable projective Pj , j = i, to Pj and the

indecomposable projective Pi to the cone over the morphism

Pi → Pj .
arrows of Q
i→j
In fact, the sum of the images of all the Pj has a natural structure of complex of
kQ -kQ-bimodules and F can also be described as the derived tensor product
over kQ with this complex of bimodules.
The example of the following quiver Q
1
E 222 a
b 22

2o c 3
and its mutation Q at 1
Q : 2 ← 1 ← 3
shows that if we mutate at a vertex which is neither a sink nor a source, then
the derived categories of representations of Q and Q are not equivalent in
general. The cluster-tilted algebra associated with the mutation of Q at 1 (cf.

section 5.8) is the quotient of kQ by the ideal generated by ab, bc and ca. It is
of infinite global dimension and therefore not derived equivalent to kQ, either.
Clearly, in order to understand mutation from a representation-theoretic
point of view, more structure is needed. Now quiver mutation has been inde-
pendently invented and investigated in the physics literature, cf. for example
equation (12.2) on page 70 in [26] (I thank to S. Fomin for this reference). In
the physics context, a crucial role is played by the so-called superpotentials, cf.
for example [18]. This lead Derksen-Weyman-Zelevinsky to their systematic
study of quivers with potentials and their mutations in [39]. We now sketch
their main result.
9.2. Mutation of quivers with potentials

Let Q be a finite quiver. Let kQ# be the completed path algebra, i.e. the com-
#
pletion of the path algebra at the ideal generated by the arrows of Q. Thus, kQ
is a topological algebra and the paths of Q form a topologial basis so that the
underlying vector space of kQ# is

kp.
p path
The continuous Hochschild homology of kQ # is the vector space HH0 obtained as

#
the quotient of kQ by the closure of the subspace generated by all commutators.
It admits a topological basis formed by the cycles of Q, i.e. the orbits of paths
p = (i|αn | . . . |α1 |i) of length n ≥ 0 with identical source and target under the
action of the cyclic group of order n. In particular, the space HH0 is a product
of copies of k indexed by the vertices if Q does not have oriented cycles. For
each arrow a of Q, the cyclic derivative with respect to a is the unique linear
map
#
∂a : HH0 → kQ
which takes the class of a path p to the sum

vu
p=uav
taken over all decompositions of p as a concatenation of paths u, a, v, where

u and v are of length ≥ 0. A potential on Q is an element W of HH0 whose
expansion in the basis of cycles does not involve cycles of length 0.
Now assume that Q does not have loops or 2-cycles.
146 Bernhard Keller
Theorem 9.1 (Derksen-Weyman-Zelevinsky [39]). The mutation operation
Q
→ μi (Q)
admits a good extension to quivers with potentials
(Q, W )
→ μi (Q, W ) = (Q , W ) ,
i.e. the quiver Q is isomorphic to μi (Q) if W is generic.
Here, ‘generic’ means that W avoids a certain countable union of hypersur-

faces in the space of potentials. The main ingredient of the proof of this theorem
is the construction of a ‘minimal model’, which, in a different language, was
also obtained in [82] and [97].
If W is not generic, then μi (Q) is not necessarily isomorphic to Q but still
isomorphic to its 2-reduction, i.e. the quiver obtained from Q by removing
the arrows of a maximal set of pairwise disjoint 2-cycles. For example, the
mutation of the quiver
1
Q: E 222 a
b (9.2.1)
22

2o c 3
with the potential W = abc at the vertex 1 is the quiver with potential
Q : 2 ← 1 ← 3 , W = 0.
On the other hand, the mutation of the above cyclic quiver Q with the potential
W = abcabc at the vertex 1 is the quiver
1 ^=
b ===a
==
=
o
e /
2 3
c
with the potential ecec + eb a .

Two quivers with potentials (Q, W ) and (Q , W ) are right equivalent if
there is an isomorphism ϕ : kQ# → kQ $ taking W to W . It is shown in [39]
(cf. also [133]) that the mutation μi induces an involution on the set of right
equivalence classes of quivers with potentials, where the quiver does not have
loops and does not have a 2-cycle passing through i.
9.3. Derived equivalence of Ginzburg dg algebras

We will associate a derived equivalence of differential graded (=dg) algebras
with each mutation of a quiver with potential. The dg algebra in question is the
algebra = (Q, W ) associated by Ginzburg with an arbitrary quiver Q with
potential W , cf. section 4.3 of [71]. The underlying graded algebra of (Q, W )
is the (completed) graded path algebra of a graded quiver, i.e. a quiver where
each arrow has an associated integer degree. We first describe this graded quiver
It has the same vertices as Q. Its arrows are
Q:
r the arrows of Q (they all have degree 0),

r an arrow a ∗ : j → i of degree −1 for each arrow a : i → j of Q,
r loops ti : i → i of degree −2 associated with each vertex i of Q.
We now consider the completion = k# (formed in the category of graded

Q
algebras) and endow it with the unique continuous differential of degree 1 such
that on the generators we have:
r da = 0 for each arrow a of Q,

r d(a ∗ ) = ∂a W for each arrow a of Q,
r d(ti ) = ei ( [a, a ∗ ])ei for each vertex i of Q, where ei is the idempotent
a
associated with i and the sum runs over the set of arrows of Q.
One checks that we do have d 2 = 0 and that the homology in degree 0 of is

the Jacobi algebra as defined in [39]:
# α W | α ∈ Q1 ).
P(Q, W ) = kQ/(∂
Let us consider two typical examples: Let Q be the quiver with one ver-
tex and three loops labeled X, Y and Z. Let W = XY Z − XZY . Then the
cyclic derivatives of W yield the commutativity relations between X, Y and Z
and the Jacobi algebra is canonically isomorphic to the power series algebra
k[[X, Y, Z]]. It is not hard to check that in this example, the homology of is
concentrated in degree 0 so that we have a quasi-isomorphism → P(Q, W ).
Using theorem 5.3.1 of Ginzburg’s [71], one can show that this is the case if and
only if the full subcategory of the derived category of the Jacobi algebra formed
by the complexes whose homology is of finite total dimension is 3-Calabi-Yau
as a triangulated category (cf. also below).
As a second example, we consider the Ginzburg dg algebra associated
with the cyclic quiver 9.2.1 with the potential W = abc. Here is the graded
148 Bernhard Keller

quiver Q:
t2

@ 2 =^==== ∗
b∗ ====a

==
a ===
b
:1o
c
t1 /3d t3
c∗
The differential is given by
d(a ∗ ) = bc , d(b∗ ) = ca , d(c∗ ) = ab , d(t1 ) = cc∗ − b∗ b , . . . .
It is not hard to show that for each i ≤ 0, we have a canonical isomorphism
H i () = CA3 (T , i T )
where A3 is the quiver 1 → 2 → 3, CA3 its cluster category and T the sum of the
images of the modules P1 , P3 and P3 /P2 . Since is an autoequivalence of finite
order of the cluster category CA3 , we see that has non vanishing homology
in infinitely many degrees < 0. In particular, is not quasi-isomorphic to the
Jacobi algebra.
Let us denote by D the derived category of . Its objects are the differential
Z-graded right -modules and its morphisms obtained from morphisms of dg
-modules (homogeneous of degree 0 and which commute with the differential)
by formally inverting all quasi-isomorphisms, cf. [89]. Let us denote by per
the perfect derived category, i.e. the full subcategory of D which is the closure
of the free right -module . under shifts, extensions and passage to direct
factors. Finally, we denote by Df d () the finite-dimensional derived category,
i.e. the full subcategory of D formed by the dg modules whose homology
is of finite total dimension (!). We recall that the objects of per() can be
intrinsically characterized as the compact ones, i.e. those whose associated
covariant Hom-functor commutes with arbitrary coproducts. The objects M of
the bounded derived category are characterized by the fact that Hom(P , M) is
finite-dimensional for each object P of per().
The following facts will be shown in [85]. Notice that they hold for arbitrary
Q and W .
1) The dg algebra is homologically smooth, i.e. it is perfect as an object
of the derived category of e = ⊗ op . This implies that the bounded derived
category is contained in the perfect derived category, cf. e.g. [88].
2) The dg algebra is 3-Calabi-Yau as a bimodule, i.e. there is an isomor-

phism in the derived category of e
∼
RHome (, e ) → [−3].
As a consequence, the finite-dimensional derived category Df d () is 3-Calabi-

Yau as a triangulated category, cf. e.g. lemma 4.1 of [88].
3) The triangulated category D() admits a t-structure whose left aisle
D≤0 consists of the dg modules M such that H i (M) = 0 for all i > 0. This
t-structure induces a t-structure on Df d () whose heart is equivalent to the
category of finite-dimensional (and hence nilpotent) modules over the Jacobi
algebra.
The generalized cluster category [2] is the idempotent completion C(Q,W ) of
the quotient
per()/Df d ().
The name is justified by the fact that if Q is a quiver without oriented cycles
(and so W = 0), then C(Q,W ) is triangle equivalent to CQ . In general, the cate-
gory C(Q,W ) has infinite-dimensional Hom-spaces. However, if H 0 () is finite-
dimensional, then C(Q,W ) is a 2-Calabi-Yau category with cluster-tilting object
T = , in the sense of section 7.1, cf. [2] (the present version of [2] uses non
completed Ginzburg algebras; the complete case will be included in a future
version).
From now on, we suppose that W does not involve cycles of length ≤ 1.

Then the simple Q-modules Si associated with the vertices of Q yield a basis
of the Grothendieck group K0 (Df d ()). Thanks to the Euler form
P , M = χ (RHom(P , M)) , P ∈ per() , M ∈ Df d () ,
this Grothendieck group is dual to K0 (per()) and the dg modules Pi = ei
form the basis dual to the Si . The Euler form also induces an antisymmetric
bilinear form on K0 (Df d () (‘Poisson form’). If W does not involve cycles of
length ≤ 2, then the quiver of the Jacobi algebra is Q and the matrix of the
form on Df d () in the basis of the Si is given by
Si , Sj = |{arrows j → i of Q}| − |{arrows i → j of Q}.
The dual of this form is given by the map (‘symplectic form’)
op ) = K0 (per()) ⊗Z K0 (per())
K0 (k) → K0 (per( ⊗
associated with the k- e -bimodule .
Now suppose that W does not involve cycles of length ≤ 2 and that Q does
not have loops nor 2-cycles. Then each simple Si is a (3-)spherical object in the
150 Bernhard Keller
3-Calabi-Yau category Df d (), i.e. we have an isomorphism of graded algebras

Ext∗ (Si , Si ) → H ∗ (S 3 , k).
∼
Moreover, the quiver Q encodes the dimensions of the Ext1 -groups between
the Si : We have
dim Ext1 (Si , Sj ) = |{arrows j → i in Q}|.
In particular, for i = j , we have either Ext1 (Si , Sj ) = 0 or Ext1 (Sj , Si ) = 0
because Q does not contain 2-cycles. It follows that the Si , i ∈ Q0 , form a
spherical collection in Df d () in the sense of [98]. Conversely, each spherical
collection in an (A∞ -) 3-Calabi-Yau category in the sense of [98] can be
obtained in this way from the Ginzburg algebra associated to a quiver with
potential.
The categories we have considered so far are summed up in the sequence of
0 / Df d () / per() / C(Q,W ) /0. (9.3.1)

This sequence is ‘exact up to factors’, i.e. the left hand term is a thick sub-
category of the middle term and the right hand term is the idempotent closure
of the quotient. Notice that the left hand category is 3-Calabi-Yau, the middle
one does not have a Serre functor and the right hand one is 2-Calabi-Yau if the
Jacobi algebra is finite-dimensional.
We keep the last hypotheses: Q does not have loops or 2-cycles and W does
not involve cycles of length ≤ 2. Let i be a vertex of Q and the Ginzburg
algebra of the mutated quiver with potential μi (Q, W ). The following theorem
improves on Vitória’s [126], cf. also [107] [18].
Theorem 9.2 ([94]). There is an equivalence of derived categories
F : D( ) → D()
which takes the dg modules Pj to Pj for j = i and Pi to the cone on the
morphism

Pi −→ Pj .
arrows
i→j
It induces triangle equivalences per( ) → per() and Df d ( ) → Df d ().

∼
In fact, the functor F is given by the left derived functor of the tensor product
by a suitable --bimodule.
By transport of the canonical t-structure on D( ), we obtain new t-structures
on D() and Df d (). They are related to the canonical ones by a tilt (in the
sense of [75]) at the simple object corresponding to the vertex i. By iterating

mutation (as far as possible), one obtains many new t-structures on Df d ().
In the context of Iyama-Reiten’s [80], the dg algebras and have their
homologies concentrated in degree 0 and the equivalence of the theorem is
given by the tilting module constructed in [loc. cit.].
9.4. A geometric illustration

To illustrate the sequence 9.3.1, let us consider the example of the quiver
@ @ 2 =
========
@ ======
Q: =====
=

0 oo 1
o
where the arrows going out from i are labeled xi , yi , zi , 0 ≤ i ≤ 2, endowed
with the potential

2
W = (xi yi zi − xi zi yi ).
i=0
Let p : ω → P2 be the canonical bundle on P2 . Then we have a triangle equiv-

alence
∼
Db (coh(ω)) → per()
which sends p ∗ (O(−i)) to the dg module Pi , 0 ≤ i ≤ 2. Under this equiva-

lence, the subcategory Df d corresponds to the subcategory DZb (coh(ω)) of
complexes of coherent sheaves whose homology is supported on the zero sec-
tion Z of the bundle ω. This subcategory is indeed Calabi-Yau of dimension
3. Bridgeland has studied its t-structures [22] by linking them to t-structures
and mutations (in the sense of Rudakov’s school) in the derived category of
coherent sheaves on the projective plane. In this example, the Ginzburg algebra
has its homology concentrated in degree 0. So it is quasi-isomorphic to the
Jacobi algebra. The Jacobi algebra is infinite-dimensional so that the general-
ized cluster category C(Q,W ) is not Hom-finite. The category C(Q,W ) identifies
with the quotient
Db (coh(ω))/DZb (coh(ω))
and thus with the bounded derived category Db (coh(ω \ Z)) of coherent
sheaves on the complement of the zero section of ω. In order to understand
why this category is ‘close to being’ Calabi-Yau of dimension 2, we consider
152 Bernhard Keller
the scheme C obtained from ω by contracting the zero section Z to a point P0 .

The projection ω → C induces an isomorphism from ω \ Z onto C \ {P0 } and
so we have an equivalence
∼ ∼
C(Q,W ) → Db (coh(ω \ Z)) → Db (coh(C \ {P0 })).
Let C denote the completion of C at the singular point P0 . Then C is of

dimension 3 and has P0 as its unique closed point. Thus the subscheme C \ {P0 }
is of dimension 2. The induced functor
∼
\ {P0 }))
C(Q,W ) → Db (coh(C \ {P0 })) → Db (coh(C
yields a ‘completion’ of C(Q,W ) which is of ‘dimension 2’ and Calabi-Yau in a
generalized sense, cf. [35].
9.5. Ginzburg algebras from algebras of global dimension 2

Let us link the cluster categories obtained from algebras of global dimen-
sion 2 in section 7.7 to the generalized cluster categories C(Q,W ) constructed
above.
Let A be an algebra given as the quotient kQ /I of the path algebra of a finite
quiver Q by an ideal I contained in the square of the ideal J generated by the
arrows of Q . Assume that A is of global dimension 2 (but not necessarily of
finite dimension over k). We construct a quiver with potential (Q, W ) as follows:
Let R be the union over all pairs of vertices (i, j ) of a set of representatives of
the vectors belonging to a basis of
op
2 (Sj , Si ) = ej (I /(I J + J I ))ei .
TorA
We think of these representatives as ‘minimal relations’ from i to j , cf. [20].

For each such representative r, let ρr be a new arrow from j to i. We define Q
to be obtained from Q by adding all the arrows ρr and the potential is given
by

W = rρr .
r∈R
Recall that a tilting module over an algebra B is a B-module T such that the
total derived functor of the tensor product by T over the endomorphism algebra
EndB (T ) is an equivalence
∼
D(EndB (T )) → D(B).
The second assertion of part a) of the following theorem generalizes a result
by Assem-Brüstle-Schiffler [3].
Theorem 9.3 ([85]).

a) The category C(Q,W ) is triangle equivalent to the cluster category CA . This
equivalence takes to the image π(A) of A in CA and thus induces an
isomorphism from the Jacobi algebra P(Q, W ) onto the endomorphism
algebra of the image of A in CA .
b) If T is a tilting module over kQ for a quiver without oriented cycles Q
and A is the endomorphism algebra of T , then C(Q,W ) is triangle equivalent
to CQ .
9.6. Cluster-tilting objects, spherical collections,

decorated representations
We consider the setup of theorem 9.2. The equivalence F induces equivalences
per( ) → per() and Df d ( ) → Df d (). The last equivalence takes Si to
Si and, for j = i, the module Sj to Sj if Ext1 (Sj , Si ) = 0 and, more generally,
to the middle term F Sj of the universal extension
Sid → F Sj → Sj → Sid ,

where the components of the third morphism form a basis of Ext1 (Sj , Si ). This
means that the images F Sj , j ∈ Q0 , form the left mutated spherical collection
in Df d () in the sense of [98].
If H 0 () is finite-dimensional so that C(Q,W ) is a 2-Calabi-Yau category with
cluster-tilting object, then the images of the F Pj , j ∈ Q0 , form the mutated
cluster-tilting object, as it follows from the description in lemma 7.5.
Now suppose that H 0 () is finite-dimensional. Then we can establish a
connection with decorated representations and their mutations in the sense of
[39]: Recall from [loc. cit.] that a decorated representation of (Q, W ) is given
by a finite-dimensional (hence nilpotent) module M over the Jacobi algebra
and a collection of vector spaces Vj indexed by the vertices j of Q. Given an
object L of C(Q,W ) , we put M = C(, L) and, for each vertex j , we choose a
vector space Vj of maximal dimension such that the triangle
T1 → T0 → L → T1 (9.6.1)
of lemma 7.8 admits a direct factor
Vj ⊗ Pj → 0 → Vj ⊗ Pj → Vj ⊗ Pj .
Let us write GL for the decorated representation thus constructed. The assign-
ment L
→ GL defines a bijection between isomorphism classes of objects L
in C(Q,W ) and right equivalence classes of decorated representations. It is com-
patible with mutations: If (M , V ) is the image GL of an object L of C(Q ,W ) ,
154 Bernhard Keller
then the mutation (Q, W, M, V ) of (Q , W , M , V ) at i in the sense of [39]

is right equivalent to (Q, W, M , V ) where (M , V ) = GF L and F is the
equivalence of theorem 9.2.
References
[1] Claire Amiot, Cluster categories for algebras of global dimension 2 and quivers
with potential, Annales de l’institut Fourier 59 (2009), no. 6, 2525–2590.
[2] Claire Amiot, Sur les petites catégories triangulées, Ph. D. thesis, Université
Paris Diderot - Paris 7, Juillet 2008.
[3] Ibrahim Assem, Thomas Brüstle, and Ralf Schiffler, Cluster-tilted algebras as
trivial extensions, Bull. London Math. Soc. 40 (2008), 151–162.
[4] Ibrahim Assem, Daniel Simson, and Andrzej Skowroński, Elements of the rep-
resentation theory of associative algebras. Vol. 1, London Mathematical Society
Student Texts, vol. 65, Cambridge University Press, Cambridge, 2006, Tech-
niques of representation theory.
[5] M. Auslander, I. Reiten, and S. Smalø, Representation theory of Artin algebras,
Cambridge Studies in Advanced Mathematics, vol. 36, Cambridge University
Press, 1995 (English).
[6] Maurice Auslander, Functors and morphisms determined by objects, Represen-
tation theory of algebras (Proc. Conf., Temple Univ., Philadelphia, Pa., 1976),
Dekker, New York, 1978, pp. 1–244. Lecture Notes in Pure Appl. Math., Vol. 37.
[7] Aslak Bakke Buan, Osamu Iyama, Idun Reiten, and Jeanne Scott,
Cluster structures for 2-Calabi-Yau categories and unipotent groups,
arXiv:math.RT/0701557. Compos. Math. 145 (2009), no. 4, 1035–1079.
[8] Aslak Bakke Buan, Osamu Iyama, Idun Reiten, and David Smith, Mutation of
cluster-tilting objects and potentials, arXiv:0804.3813. to appear in Amer. J.
Math.
[9] Aslak Bakke Buan and Robert Marsh, Cluster-tilting theory, Trends in represen-
tation theory of algebras and related topics, Contemp. Math., vol. 406, Amer.
Math. Soc., Providence, RI, 2006, pp. 1–30.
[10] Aslak Bakke Buan, Robert J. Marsh, Markus Reineke, Idun Reiten, and Gordana
Todorov, Tilting theory and cluster combinatorics, Advances in Mathematics 204
(2) (2006), 572–618.
[11] Aslak Bakke Buan, Robert J. Marsh, and Idun Reiten, Cluster mutation via quiver
representations, Comm. Math. Helv. 83 (2008), 143–177.
[12] Aslak Bakke Buan, Robert J. Marsh, Idun Reiten, and Gordana Todorov, Clusters
and seeds in acyclic cluster algebras, Proc. Amer. Math. Soc. 135 (2007), no. 10,
3049–3060 (electronic), With an appendix coauthored in addition by P. Caldero
and B. Keller.
[13] Aslak Bakke Buan and Idun Reiten, Acyclic quivers of finite mutation type, Int.
Math. Res. Not. (2006), Art. ID 12804, 10.
[14] Michael Barot, Christof Geiss, and Andrei Zelevinsky, Cluster algebras of finite
type and positive symmetrizable matrices, J. London Math. Soc. (2) 73 (2006),
no. 3, 545–564.
[15] Andre Beineke, Thomas Bruestle, and Lutz Hille, Cluster-cyclic quivers with
three vertices and the Markov equation, arXiv:math/0612213. to appear in Algebr.
Represent. Theory.
[16] Arkady Berenstein, Sergey Fomin, and Andrei Zelevinsky, Parametrizations of
canonical bases and totally positive matrices, Adv. Math. 122 (1996), no. 1,
49–149.
[17] , Cluster algebras. III. Upper bounds and double Bruhat cells, Duke
Math. J. 126 (2005), no. 1, 1–52.
[18] David Berenstein and Michael Douglas, Seiberg duality for quiver gauge theories,
arXiv:hep-th/0207027v1.
[19] I. N. Bernšteı̆n, I. M. Gel fand, and V. A. Ponomarev, Coxeter functors, and
Gabriel’s theorem, Uspehi Mat. Nauk 28 (1973), no. 2(170), 19–33.
[20] Klaus Bongartz, Algebras and quadratic forms, J. London Math. Soc. (2) 28
(1983), no. 3, 461–469.
[21] Tom Bridgeland, Stability conditions and Hall algebras, Talk at the meeting
‘Recent developments in Hall algebras’, Luminy, November 2006.
[22] , t-structures on some local Calabi-Yau varieties, J. Algebra 289 (2005),
no. 2, 453–483.
[23] , Derived categories of coherent sheaves, International Congress of Math-
ematicians. Vol. II, Eur. Math. Soc., Zürich, 2006, pp. 563–582.
[24] , Stability conditions on triangulated categories, Ann. of Math. (2) 166
(2007), no. 2, 317–345.
[25] Igor Burban, Osamu Iyama, Bernhard Keller, and Idun Reiten, Cluster tilting
for one-dimensional hypersurface singularities, Adv. Math. 217 (2008), no. 6,
2443–2484.
[26] F. Cachazo, B. Fiol, K. Intriligator, S. Katz, and C. Vafa, A geometric unification
of dualities, Nucl. Phys. B 628 (2002), 3–78.
[27] Philippe Caldero and Frédéric Chapoton, Cluster algebras as Hall algebras of
quiver representations, Comment. Math. Helv. 81 (2006), no. 3, 595–616.
[28] Philippe Caldero, Frédéric Chapoton, and Ralf Schiffler, Quivers with relations
arising from clusters (An case), Trans. Amer. Math. Soc. 358 (2006), no. 5,
1347–1364.
[29] Philippe Caldero and Bernhard Keller, From triangulated categories to cluster
algebras. II, Ann. Sci. École Norm. Sup. (4) 39 (2006), no. 6, 983–1009.
[30] Philippe Caldero and Bernhard Keller, From triangulated categories to cluster
algebras, Inv. Math. 172 (2008), 169–211.
[31] Philippe Caldero and Markus Reineke, On the quiver Grassmannian in the acyclic
case, Journal of Pure and Applied Algebra. 212 (2008), no. 11, 2369–2380.
[32] Henri Cartan and Samuel Eilenberg, Homological algebra, Princeton University
Press, Princeton, N. J., 1956.
[33] Frédéric Chapoton, Enumerative properties of generalized associahedra, Sém.
Lothar. Combin. 51 (2004/05), Art. B51b, 16 pp. (electronic).
[34] Frédéric Chapoton, Sergey Fomin, and Andrei Zelevinsky, Polytopal realizations
of generalized associahedra, Canad. Math. Bull. 45 (2002), no. 4, 537–566,
Dedicated to Robert V. Moody.
[35] Joe Chuang and Raphaël Rouquier, book in preparation.
156 Bernhard Keller
[36] William Crawley-Boevey, On the exceptional fibres of Kleinian singularities,

Amer. J. Math. 122 (2000), no. 5, 1027–1037.
[37] Raika Dehy and Bernhard Keller, On the combinatorics of rigid objects in
2-Calabi-Yau categories, International Mathematics Research Notices 2008
(2008), rnn029–17.
[38] Harm Derksen and Theodore Owen, New graphs of finite mutation type, Electron.
J. Combin. 15 (2008), no. 1. Research Paper 139, 15pp.
[39] Harm Derksen, Jerzy Weyman, and Andrei Zelevinsky, Quivers with potentials
and their representations I: Mutations, Selecta Mathematica. 14 (2008), 59–119.
[40] Vlastimil Dlab and Claus Michael Ringel, Indecomposable representations of
graphs and algebras, Mem. Amer. Math. Soc. 6 (1976), no. 173, v+57.
[41] Peter Donovan and Mary Ruth Freislich, The representation theory of finite
graphs and associated algebras, Carleton University, Ottawa, Ont., 1973, Car-
leton Mathematical Lecture Notes, No. 5.
[42] M. Fekete, Ueber ein Problem von Laguerre, Rend. Circ. Mat. Palermo 34 (1912),
89–100, 110–120.
[43] V. V. Fock and A. B. Goncharov, Cluster ensembles, quantization and the dilog-
arithm, Annales scientifiques de l’ENS 42 (2009), no. 6, 865–930.
[44] , Cluster ensembles, quantization and the dilogarithm II: The intertwiner,
arXiv:math/0702398.
[45] , The quantum dilogarithm and representations of quantized cluster vari-
eties, arXiv:math/0702397. to appear in Jnuent. Math.
[46] V. V. Fock and A. B. Goncharov, Cluster X -varieties, amalgamation, and
Poisson-Lie groups, Algebraic geometry and number theory, Progr. Math., vol.
253, Birkhäuser Boston, Boston, MA, 2006, pp. 27–68.
[47] Sergey Fomin, Cluster algebras portal,

.
[48] Sergey Fomin and Nathan Reading, Generalized cluster complexes and Coxeter
combinatorics, Int. Math. Res. Not. (2005), no. 44, 2709–2757.
[49] Sergey Fomin, Michael Shapiro, and Dylan Thurston, Cluster algebras and
triangulated surfaces. Part I: Cluster complexes, Acta Mathematica. 201 (2008),
no. 1, 83–146.
[50] Sergey Fomin and Andrei Zelevinsky, Cluster algebras. I. Foundations, J. Amer.
Math. Soc. 15 (2002), no. 2, 497–529 (electronic).
[51] , The Laurent phenomenon, Adv. in Appl. Math. 28 (2002), no. 2, 119–
144.
[52] , Cluster algebras. II. Finite type classification, Invent. Math. 154 (2003),
no. 1, 63–121.
[53] , Cluster algebras: notes for the CDM-03 conference, Current develop-
ments in mathematics, 2003, Int. Press, Somerville, MA, 2003, pp. 1–34.
[54] , Y -systems and generalized associahedra, Ann. of Math. (2) 158 (2003),
no. 3, 977–1018.
[55] , Cluster algebras IV: Coefficients, Compositio Mathematica 143 (2007),
112–164.
[56] Edward Frenkel and András Szenes, Thermodynamic Bethe ansatz and diloga-
rithm identities. I, Math. Res. Lett. 2 (1995), no. 6, 677–693.
[57] Changjian Fu and Bernhard Keller, On cluster algebras with coefficients and
2-Calabi-Yau categories, Trans. Amer. Math. Soc. 362 (2010), 859–895.
[58] P. Gabriel, Unzerlegbare Darstellungen I, Manuscripta Math. 6 (1972), 71–103.
[59] P. Gabriel and A.V. Roiter, Representations of finite-dimensional algebras, Ency-
clopaedia Math. Sci., vol. 73, Springer–Verlag, 1992.
[60] Peter Gabriel, Auslander-Reiten sequences and representation-finite algebras,
Representation theory, I (Proc. Workshop, Carleton Univ., Ottawa, Ont., 1979),
Springer, Berlin, 1980, pp. 1–71.
[61] Christof Geiß, Bernard Leclerc, and Jan Schröer, Cluster algebra structures and
semicanonical bases for unipotent groups, arXiv:math/0703039.
[62] , Partial flag varieties and preprojective algebras, Ann. Jnst. Fourier.
(Grenoble) 58 (2008), no. 3, 825–876.
[63] , Preprojective algebras and cluster algebras, Trends in representation
theory and related topics, EMS Ser. Congr. Rep., Eur. Math. Soc., Zürich 2008,
pp. 253–283.
[64] , Semicanonical bases and preprojective algebras, Ann. Sci. École Norm.
Sup. (4) 38 (2005), no. 2, 193–253.
[65] , Rigid modules over preprojective algebras, Invent. Math. 165 (2006),
no. 3, 589–632.
[66] , Auslander algebras and initial seeds for cluster algebras, J. Lond. Math.
Soc. (2) 75 (2007), no. 3, 718–740.
[67] , Semicanonical bases and preprojective algebras. II. A multiplication
formula, Compos. Math. 143 (2007), no. 5, 1313–1334.
[68] Michael Gekhtman, Michael Shapiro, and Alek Vainshtein, On the properties
of the exchange graph of a cluster algebra, Math. Res. Lett. 15 (2008), no. 2,
321–330.
[69] , Cluster algebras and Poisson geometry, Mosc. Math. J. 3 (2003), no. 3,
899–934, 1199, {Dedicated to Vladimir Igorevich Arnold on the occasion of his
65th birthday}.
[70] , Cluster algebras and Weil-Petersson forms, Duke Math. J. 127 (2005),
no. 2, 291–311.
[71] Victor Ginzburg, Calabi-Yau algebras, arXiv:math/0612139v3 [math.AG].
[72] F. Gliozzi and R. Tateo, Thermodynamic Bethe ansatz and three-fold triangula-
tions, Internat. J. Modern Phys. A 11 (1996), no. 22, 4051–4064.
[73] Dieter Happel, On the derived category of a finite-dimensional algebra, Com-
ment. Math. Helv. 62 (1987), no. 3, 339–389.
[74] , Triangulated categories in the representation theory of finite-
dimensional algebras, Cambridge University Press, Cambridge, 1988.
[75] Dieter Happel, Idun Reiten, and Sverre O. Smalø, Tilting in abelian categories
and quasitilted algebras, Mem. Amer. Math. Soc. 120 (1996), no. 575, viii+ 88.
[76] Dieter Happel and Luise Unger, On a partial order of tilting modules, Algebr.
Represent. Theory 8 (2005), no. 2, 147–156.
[77] André Henriques, A periodicity theorem for the octahedron recurrence, J. Alge-
braic Combin. 26 (2007), no. 1, 1–26.
[78] David Hernandez and Bernard Leclerc, Cluster algebras and quantum affine
algebras, arXiv:0903.1452v1 [math.Q4].
158 Bernhard Keller
[79] Andrew Hubery, Acyclic cluster algebras via Ringel-Hall algebras, Preprint
available at the author’s home page.
[80] Osamu Iyama and Idun Reiten, Fomin-Zelevinsky mutation and tilting modules
over Calabi-Yau algebras, Amer. J. Math. 130 (2008), no. 4, 1087–1149.
[81] Osamu Iyama and Yuji Yoshino, Mutations in triangulated categories and rigid
Cohen-Macaulay modules, Inv. Math. 172 (2008), 117–168.
[82] Hiroshige Kajiura, Noncommutative homotopy algebras associated with open
strings, Rev. Math. Phys. 19 (2007), no. 1, 1–99.
[83] Masaki Kashiwara, Bases cristallines, C. R. Acad. Sci. Paris Sér. I Math. 311
(1990), no. 6, 277–280.
[84] Masaki Kashiwara and Pierre Schapira, Sheaves on manifolds, Grundlehren der
Mathematischen Wissenschaften [Fundamental Principles of Mathematical Sci-
ences], vol. 292, Springer-Verlag, Berlin, 1994, With a chapter in French by
Christian Houzel, Corrected reprint of the 1990 original.
[85] Bernhard Keller, Deformed CY-completions, arXiv:0908.3499 [math.RT].
[86] , The periodicity conjecture for pairs of Dynkin diagrams,
arXiv:1001:1531.
[87] , Quiver mutation in Java, Java applet available at the author’s home
page.
[88] , Triangulated Calabi-Yau categories, Trends in Representation Theory
of Algebras (A. Skowronski, ed.), Eur. Math. Soc., 2008, pp. 467–489.
[89] , Deriving DG categories, Ann. Sci. École Norm. Sup. (4) 27 (1994),
no. 1, 63–102.
[90] , On triangulated orbit categories, Doc. Math. 10 (2005), 551–581.
[91] Bernhard Keller and Idun Reiten, Acyclic Calabi-Yau categories are clus-
ter categories, preprint, 2006, with an appendix by Michel Van den Bergh,
arXiv:math.RT/0610594, to appear in Compos. Math.
[92] , Cluster-tilted algebras are Gorenstein and stably Calabi-Yau, Advances
in Mathematics 211 (2007), 123–151.
[93] Bernhard Keller and Michel Van den Bergh, On two examples of Iyama and
Yoshino, arXiv:0803.0720.
[94] Bernhard Keller and Dong Yang, Derived equivalences from mutations of quivers
with potential, arXiv:0906:0761.
[95] Maxim Kontsevich, Donaldson-Thomas invariants, Mathematische Arbeitsta-
gung June 22-28, 2007, MPI Bonn, www.mpim-bonn.mpg.de/preprints/.
[96] , Donaldson-Thomas invariants, stability conditions and cluster trans-
formations, Report on joint work with Y. Soibelman, talks at the IHES, October
11 and 12, 2007.
[97] Maxim Kontsevich and Yan Soibelman, Notes on A-infinity algebras, A-infinity
categories and non-commutative geometry. I, arXiv:math.RA/0606241.
[98] , Stability structures, Donaldson-Thomas invariants and cluster trans-
formations, arXiv:0811.2435.
[99] Christian Krattenthaler, The F -triangle of the generalised cluster complex, Topics
in discrete mathematics, Algorithms Combin., vol. 26, Springer, Berlin, 2006,
pp. 93–126.
[100] A. Kuniba and T. Nakanishi, Spectra in conformal field theories from the Rogers
dilogarithm, Modern Phys. Lett. A 7 (1992), no. 37, 3487–3494.
[101] Atsuo Kuniba, Tomoki Nakanishi, and Junji Suzuki, Functional relations in solv-
able lattice models. I. Functional relations and representation theory, Internat.
J. Modern Phys. A 9 (1994), no. 30, 5215–5266.
[102] G. Lusztig, Canonical bases arising from quantized enveloping algebras, J. Amer.
Math. Soc. 3 (1990), no. 2, 447–498.
[103] , Total positivity in reductive groups, Lie theory and geometry, Progr.
Math., vol. 123, Birkhäuser Boston, Boston, MA, 1994, pp. 531–568.
[104] , Semicanonical bases arising from enveloping algebras, Adv. Math. 151
(2000), no. 2, 129–139.
[105] George Lusztig, Canonical bases and Hall algebras, Representation theories and
algebraic geometry (Montreal, PQ, 1997), NATO Adv. Sci. Inst. Ser. C Math.
Phys. Sci., vol. 514, Kluwer Acad. Publ., Dordrecht, 1998, pp. 365–399.
[106] Robert Marsh, Markus Reineke, and Andrei Zelevinsky, Generalized associa-
hedra via quiver representations, Trans. Amer. Math. Soc. 355 (2003), no. 10,
4171–4186 (electronic).
[107] Subir Mukhopadhyay and Koushik Ray, Seiberg duality as derived equivalence
for some quiver gauge theories, J. High Energy Phys. (2004), no. 2, 070, 22 pp.
(electronic).
[108] Gregg Musiker, A graph theoretic expansion formula for cluster algebras of
type Bn and Dn , arXiv:0710.3574v1 [math.CO]. to appear in the Annals of
Combinatorics.
[109] Hiraku Nakajima, Quiver varieties and cluster algebras, arXiv:0905002
[math.QA].
[110] L. A. Nazarova, Representations of quivers of infinite type, Izv. Akad. Nauk SSSR
Ser. Mat. 37 (1973), 752–791.
[111] Yann Palu, Grothendieck group and generalized mutation rule for 2-calabi-yau
triangulated categories, Journal of Pure and Applied Algebra. 213 (2009), no. 7,
1438–1449.
[112] , Cluster characters for 2-calabi-yau triangulated categories, Annales
de l’institut Fourier 58 (2008), no. 6, 2221–2248.
[113] F. Ravanini, A. Valleriani, and R. Tateo, Dynkin TBAs, Internat. J. Modern Phys.
A 8 (1993), no. 10, 1707–1727.
[114] Idun Reiten, Tilting theory and cluster algebras, preprint available at
www.institut.math.jussieu.fr/keller/ictp2006/lecturenotes/reiten.pdf.
[115] Claus Michael Ringel, The preprojective algebra of a quiver, Algebras and
modules, II (Geiranger, 1996), CMS Conf. Proc., vol. 24, Amer. Math. Soc.,
Providence, RI, 1998, pp. 467–480.
[116] , Some remarks concerning tilting modules and tilted algebras. Origin.
Relevance. Future., Handbook of Tilting Theory, LMS Lecture Note Series, vol.
332, Cambridge Univ. Press, Cambridge, 2007, pp. 49–104.
[117] C.M. Ringel, Tame algebras and integral quadratic forms, Lecture Notes in
Mathematics, vol. 1099, Springer Verlag, 1984.
[118] Kyoji Saito, Extended affine root systems. I. Coxeter transformations, Publ. Res.
Inst. Math. Sci. 21 (1985), no. 1, 75–179.
[119] Olivier Schiffmann, Lectures on Hall algebras, math.RT/0611617.
[120] Joshua S. Scott, Grassmannians and cluster algebras, Proc. London Math. Soc.
(3) 92 (2006), no. 2, 345–380.
160 Bernhard Keller
[121] Catherine Soanes and Angus Stevenson, Concise Oxford English Dictionary,
Oxford University Press, Oxford, 2004.
[122] James Dillon Stasheff, Homotopy associativity of H -spaces. I, II, Trans. Amer.
Math. Soc. 108 (1963), 275–292; ibid. 108 (1963), 293–312.
[123] Balázs Szendröi, Non-commutative Donaldson-Thomas theory and the conifold,
Geom. Toplo. 12 (2008), no. 2, 1171–1202.
[124] András Szenes, Periodicity of Y -systems and flat connections, Lett. Math. Phys.
89 (2009), no. 3, 217–230.
[125] Jean-Louis Verdier, Des catégories dérivées des catégories abéliennes,
Astérisque, vol. 239, Société Mathématique de France, 1996 (French).
[126] Jorge Vitória, Mutations vs. Seiberg duality, J. Algebra 321 (2009), no. 3, 816–
828.
[127] Alexandre Yu. Volkov, On the periodicity conjecture for Y -systems, Comm. Math.
Phys. 276 (2007), no. 2, 509–517.
[128] Keiichi Watanabe, Certain invariant subrings are Gorenstein. I, II, Osaka J.
Math. 11 (1974), 1–8; ibid. 11 (1974), 379–388.
[129] Shih-Wei Yang and Andrei Zelevinsky, Cluster algebras of finite type via Coxeter
elements and principal minors, Transform. Groups 13 (2008), no. 3–4, 855–895.
[130] Yuji Yoshino, Cohen-Macaulay modules over Cohen-Macaulay rings, London
Mathematical Society Lecture Note Series, vol. 146, Cambridge University Press,
Cambridge, 1990.
[131] Al. B. Zamolodchikov, On the thermodynamic Bethe ansatz equations for reflec-
tionless ADE scattering theories, Phys. Lett. B 253 (1991), no. 3–4, 391–394.
[132] Andrei Zelevinsky, Cluster algebras: notes for 2004 IMCC (Chonju, Korea,
August 2004), arXiv:math.RT/0407414.
[133] Andrei Zelevinsky, Mutations for quivers with potentials: Oberwolfach talk, april
2007, arXiv:0706.0822.
[134] , From Littlewood-Richardson coefficients to cluster algebras in three
lectures, Symmetric functions 2001: surveys of developments and perspectives,
NATO Sci. Ser. II Math. Phys. Chem., vol. 74, Kluwer Acad. Publ., Dordrecht,
2002, pp. 253–273.
[135] , Cluster algebras: origins, results and conjectures, Advances in algebra
towards millennium problems, SAS Int. Publ., Delhi, 2005, pp. 85–105.
[136] , What is a cluster algebra?, Notices of the A.M.S. 54 (2007), no. 11,
1494–1495.
Université Paris Diderot – Paris 7, UFR de Mathématiques, Institut de

Mathématiques de Jussieu, UMR 7586 du CNRS, Case 7012, 2, place Jussieu, 75251 Paris
Cedex 05, France
E-mail address:
Localization theory for triangulated categories
henning krause
Contents
1. Introduction 161
2. Categories of fractions and localization functors 164
3. Calculus of fractions 172
4. Localization for triangulated categories 178
5. Localization via Brown representability 192
6. Well generated triangulated categories 203
7. Localization for well generated categories 213
8. Epilogue: Beyond well generatedness 224
Appendix A. The abelianization of a triangulated category 225
Appendix B. Locally presentable abelian categories 227
References 233
1. Introduction
These notes provide an introduction to the theory of localization for triangulated
categories. Localization is a machinery to formally invert morphisms in a
category. We explain this formalism in some detail and we show how it is
applied to triangulated categories.
There are basically two ways to approach the localization theory for trian-
gulated categories and both are closely related to each other. To explain this,
let us fix a triangulated category T . The first approach is Verdier localization.
For this one chooses a full triangulated subcategory S of T and constructs
a universal exact functor T → T /S which annihilates the objects belonging
to S. In fact, the quotient category T /S is obtained by formally inverting all
morphisms σ in T such that the cone of σ belongs to S.
On the other hand, there is Bousfield localization. In this case one considers
an exact functor L : T → T together with a natural morphism ηX : X → LX
for all X in T such that L(ηX) = η(LX) is invertible. There are two full
triangulated subcategories arising from such a localization functor L. We have
161
162 Henning Krause
the subcategory Ker L formed by all L-acyclic objects, and we have the essential
image Im L which coincides with the subcategory formed by all L-local objects.
Note that L, Ker L, and Im L determine each other. Moreover, L induces an
∼
equivalence T / Ker L − → Im L. Thus a Bousfield localization functor T → T
is nothing but the composite of a Verdier quotient functor T → T /S with a
fully faithful right adjoint T /S → T .
Having introduced these basic objects, there are a number of immediate
questions. For example, given a triangulated subcategory S of T , can we find
a localization functor L : T → T satisfying Ker L = S or Im L = S? On the
other hand, if we start with L, which properties of Ker L and Im L are inherited
from T ? It turns out that well generated triangulated categories in the sense of
Neeman [33] provide an excellent setting for studying these questions.
Let us discuss briefly the relevance of well generated categories. The concept
generalizes that of a compactly generated triangulated category. For example,
the derived category of unbounded chain complexes of modules over some
fixed ring is compactly generated. Also, the stable homotopy category of CW-
spectra is compactly generated. Given any localization functor L on a compactly
generated triangulated category, it is rare that Ker L or Im L are compactly
generated. However, in all known examples Ker L and Im L are well generated.
The following theorem provides a conceptual explanation; it combines several
results from Section 7.
Theorem. Let T be a well generated triangulated category and S a full
triangulated subcategory which is closed under small coproducts. Then the
following are equivalent.
(1) The triangulated category S is well generated.
(2) The triangulated category T /S is well generated.
(3) There exists a cohomological functor H : T → A into a locally presentable
abelian category such that H preserves small coproducts and S = Ker H .
(4) There exists a small set S0 of objects in S such that S admits no proper full
triangulated subcategory closed under small coproducts and containing
S0 .
Moreover, in this case there exists a localization functor L : T → T such that
Ker L = S.
Note that every abelian Grothendieck category is locally presentable; in
particular every module category is locally presentable.
Our approach for studying localization functors on well generated trian-
gulated categories is based on the interplay between triangulated and abelian
structure. A well known construction due to Freyd provides for any triangulated
category T an abelian category A(T ) together with a universal cohomological

functor T → A(T ). However, the category A(T ) is usually far too big and
therefore not manageable. If T is well generated, then we have a canonical
filtration

A(T ) = Aα (T )
α
indexed by all regular cardinals, such that for each α the category Aα (T )
is abelian and locally α-presentable in the sense of Gabriel and Ulmer [17].
Moreover, each inclusion Aα (T ) → A(T ) admits an exact right adjoint and the
composite
Hα : T −→ A(T ) −→ Aα (T )
is the universal cohomological functor into a locally α-presentable abelian cat-

egory. Thus we may think of the functors T → Aα (T ) as successive approx-
imations of T by locally presentable abelian categories. For instance, there
exists for each object X in T some cardinal α(X) such that the induced map
T (X, Y ) → Aβ (T )(Hβ X, Hβ Y ) is bijective for all Y in T and all β ≥ α(X).
These notes are organized as follows. We start off with an introduction to
categories of fractions and localization functors for arbitrary categories. Then
we apply this to triangulated categories. First we treat arbitrary triangulated cat-
egories and explain the localization in the sense of Verdier and Bousfield. Then
we pass to compactly and well generated triangulated categories where Brown
representability provides an indispensable tool for constructing localization
functors. Module categories and their derived categories are used to illustrate
most of the concepts; see [12] for complementary material from topology. The
results on well generated categories are based on facts from the theory of locally
presentable categories; we have collected these in a separate appendix.
Acknowledgement
The plan to write an introduction to the theory of triangulated localization
took shape during the “Workshop on Triangulated Categories” in Leeds 2006.
I wish to thank the organizers Thorsten Holm, Peter Jørgensen, and Raphaël
Rouquier for their skill and diligence in organizing this meeting. Most of these
notes were then written during a three months stay in 2007 at the Centre
de Recerca Matemàtica in Barcelona as a participant of the special program
“Homotopy Theory and Higher Categories”. I am grateful to the organizers
Carles Casacuberta, Joachim Kock, and Amnon Neeman for creating a stim-
ulating atmosphere and for several helpful discussions. Finally, I would like
164 Henning Krause
to thank Xiao-Wu Chen, Daniel Murfet, and Jan Šťovı́ček for their helpful
comments on a preliminary version of these notes.
2. Categories of fractions and localization functors

2.1. Categories
Throughout we fix a universe of sets in the sense of Grothendieck [19]. The
members of this universe will be called small sets.
Let C be a category. We denote by Ob C the set of objects and by Mor C
the set of morphisms in C. Given objects X, Y in C, the set of morphisms
X → Y will be denoted by C(X, Y ). The identity morphism of an object X is
denoted by idC X or just id X. If not stated otherwise, we always assume that
the morphisms between two fixed objects of a category form a small set.
A category C is called small if the isomorphism classes of objects in C
form a small set. In that case we define the cardinality of C as card C =

X,Y ∈C0 card C(X, Y ) where C0 denotes a representative set of objects of C,
meeting each isomorphism class exactly once.
Let F : I → C be a functor from a small (indexing) category I to a category
C. Then we write colim F i for the colimit of F , provided it exists. Given a
−−−→
i∈I
cardinal α, the colimit of F is called α-colimit if card I < α. An example of a
%
colimit is the coproduct i∈I Xi of a family (Xi )i∈I of objects in C where the
indexing set I is always assumed to be small. We say that a category C admits
small coproducts if for every family (Xi )i∈I of objects in C which is indexed
%
by a small set I the coproduct i∈I Xi exists in C. Analogous terminology is
used for limits and products.
2.2. Categories of fractions

Let F : C → D be a functor. We say that F makes a morphism σ of C invertible
if F σ is invertible. The set of all those morphisms which F inverts is denoted
by (F ).
Given a category C and any set of morphisms of C, we consider the
category of fractions C[ −1 ] together with a canonical quotient functor
Q : C −→ C[ −1 ]
having the following properties.
(Q1) Q makes the morphisms in invertible.

(Q2) If a functor F : C → D makes the morphisms in invertible, then there
is a unique functor F̄ : C[ −1 ] → D such that F = F̄ ◦ Q .
Note that C[ −1 ] and Q are essentially unique if they exists. Now let us
sketch the construction of C[ −1 ] and Q . At this stage, we ignore set-theoretic
issues, that is, the morphisms between two objects of C[ −1 ] need not to form
a small set. We put Ob C[ −1 ] = Ob C. To define the morphisms of C[ −1 ],
consider the quiver (i.e. oriented graph) with set of vertices Ob C and with
set of arrows the disjoint union (Mor C) −1 , where −1 = {σ −1 : Y → X |
σ : X → Y }. Let P be the set of paths in this quiver (i.e. finite sequences
of composable arrows), together with the obvious composition which is the
concatenation operation and denoted by ◦P . We define Mor C[ −1 ] as the
quotient of P modulo the following relations:
(1) β ◦P α = β ◦ α for all composable morphisms α, β ∈ Mor C.

(2) idP X = idC X for all X ∈ Ob C.
(3) σ −1 ◦P σ = idP X and σ ◦P σ −1 = idP Y for all σ : X → Y in .
The composition in P induces the composition of morphisms in C[ −1 ]. The

functor Q is the identity on objects and on Mor C the composite
inc inc can
Mor C −→ (Mor C) −1 −→ P −→ Mor C[ −1 ].
Having completed the construction of the category of fractions C[ −1 ], let

us mention that it is also called quotient category or localization of C with
respect to .
2.3. Adjoint functors

Let F : C → D and G : D → C be a pair of functors and assume that F is left
adjoint to G. We denote by
θ : F ◦ G → Id D and η : Id C → G ◦ F
the corresponding adjunction morphisms. Let = (F ) denote the set of

morphisms σ of C such that F σ is invertible. Recall that a morphism
μ : F → F between two functors is invertible if for each object X the mor-
phism μX : F X → F X is invertible.
Proposition 2.3.1. The following statements are equivalent.
(1) The functor G is fully faithful.

(2) The morphism θ : F ◦ G → Id D is invertible.
(3) The functor F̄ : C[ −1 ] → D satisfying F = F̄ ◦ Q is an equivalence.
Proof. See [18, I.1.3].

166 Henning Krause
2.4. Localization functors

A functor L : C → C is called a localization functor if there exists a morphism
η : Id C → L such that Lη : L → L2 is invertible and Lη = ηL. Note that
we only require the existence of η; the actual morphism is not part of the
definition of L. However, we will see that η is determined by L, up to a unique
isomorphism L → L.
Proposition 2.4.1. Let L : C → C be a functor and η : Id C → L be a mor-

phism. Then the following are equivalent.
(1) Lη : L → L2 is invertible and Lη = ηL.

(2) There exists a functor F : C → D and a fully faithful right adjoint G : D →
C such that L = G ◦ F and η : Id C → G ◦ F is the adjunction morphism.
Proof. (1) ⇒ (2): Let D denote the full subcategory of C formed by all objects
X such that ηX is invertible. For each X ∈ D, let θ X : LX → X be the inverse
of ηX. Define F : C → D by F X = LX and let G : D → C be the inclusion.
We claim that F and G form an adjoint pair. In fact, it is straightforward to
check that the maps
D(F X, Y ) −→ C(X, GY ), α
→ Gα ◦ ηX,
and
C(X, GY ) −→ D(F X, Y ), β
→ θ Y ◦ Fβ,
are mutually inverse bijections.

(2) ⇒ (1): Let θ : F G → Id D denote the second adjunction morphism.
Then the composites
Fη θF ηG Gθ
F −→ F GF −→ F and G −→ GF G −→ G
are identity morphisms; see [27, IV.1]. We know from Proposition 2.3.1 that
θ is invertible because G is fully faithful. Therefore Lη = GF η is invertible.
Moreover, we have
Lη = GF η = (GθF )−1 = ηGF = ηL.
Corollary 2.4.2. A functor L : C → C is a localization functor if and only if

there exists a functor F : C → D and a fully faithful right adjoint G : D → C
such that L = G ◦ F . In that case there exist a unique equivalence C[ −1 ] → D
making the following diagram commutative
C[ −1 ] WWWW
Q ggggg3
gg WWL̄WWW
ggggg WWW+
C WWWWWWW ∼
g ggggg3 C
WWWWW ggggg
F W W+ g gg G
D
where denotes the set of morphisms σ in C such that Lσ is invertible.
Proof. The characterization of a localization functor follows from Proposi-
tion 2.4.1. Now observe that equals the set of morphisms σ in C such that
F σ is invertible since G is fully faithful. Thus we can apply Proposition 2.3.1
to obtain the equivalence C[ −1 ] → D making the diagram commutative.
2.5. Local objects

Given a localization functor L : C → C, we wish to describe those objects X
∼
in C such that X − → LX. To this end, it is convenient to make the following
definition. An object X in a category C is called local with respect to a set of
morphisms if for every morphism W → W in the induced map C(W , X) →
C(W, X) is bijective. Now let F : C → D be a functor and let (F ) denote the
set of morphisms σ of C such that F σ is invertible. An object X in C is called
F -local if it is local with respect to (F ).
Lemma 2.5.1. Let F : C → D be a functor and X an object of C. Suppose there
are two morphisms η1 : X → Y1 and η2 : X → Y2 such that F ηi is invertible
and Yi is F -local for i = 1, 2. Then there exists a unique isomorphism φ : Y1 →
Y2 such that η2 = φ ◦ η1 .
Proof. The morphism η1 induces a bijection C(Y1 , Y2 ) → C(X, Y2 ) and we take
for φ the unique morphism which is sent to η2 . Exchanging the roles of η1 and
η2 , we obtain the inverse for φ.
Proposition 2.5.2. Let L : C → C be a localization functor and η : Id C → L
a morphism such that Lη is invertible. Then the following are equivalent for
an object X in C.
(1) The object X is L-local.
(2) The map C(LW, X) → C(W, X) induced by ηW is bijective for all W in C.
(3) The morphism ηX : X → LX is invertible.
(4) The map C(W, X) → C(LW, LX) induced by L is bijective for all W in C.
(5) The object X is isomorphic to LX for some object X in C.
Proof. (1) ⇒ (2): The morphism ηW belongs to (L) and therefore C(ηW, X)
is bijective if X is L-local.
168 Henning Krause
(2) ⇒ (3): Put W = X. We obtain a morphism φ : LX → X which is an

inverse for ηX. More precisely, we have φ ◦ ηX = id X. On the other hand,
ηX ◦ φ = Lφ ◦ ηLX = Lφ ◦ LηX = L(φ ◦ ηX) = id LX.
Thus ηX is invertible.
F G
(3) ⇔ (4): We use the factorization C −
→D− → C of L from Proposition 2.4.1.
Then we obtain for each W in C a factorization
∼ ∼
C(W, X) −→ C(W, LX) −→ C(F W, F X) −→ C(LW, LX)
of the map fW : C(W, X) → C(LW, LX) induced by L. Here, the first map is
induced by ηX, the second follows from the adjunction, and the third is induced
by G. Thus fW is bijective for all W iff the first map is bijective for all W iff
ηX is invertible.
(3) ⇒ (5): Take X = X.
F G
(5) ⇒ (1): We use again the factorization C − →D− → C of L from Propo-
sition 2.4.1. Fix σ in (L) and observe that F σ is invertible. Then we have
C(σ, X) ∼= C(σ, G(F X )) ∼ = D(F σ, F X ) and this implies that C(σ, X) is bijec-
tive since F σ is invertible.
Given a functor F : C → D, we denote by Im F the essential image of F ,

that is, the full subcategory of D which is formed by all objects isomorphic to
F X for some object X in C.
Corollary 2.5.3. Let L : C → C be a localization functor. Then L induces an

∼
equivalence C[(L)−1 ] − → Im L and Im L is the full subcategory of C consist-
ing of all L-local subobjects.
F G
Proof. Write L as composite C − → Im L − → C of two functors, where F X =
LX for all X in C and G is the inclusion functor. Then it follows from Corol-
∼
lary 2.4.2 that F induces an equivalence C[(L)−1 ] −
→ Im L. The second asser-
tion is an immediate consequence of Proposition 2.5.2.
Given a localization functor L : C → C and an object X in C, the morphism

X → LX is initial among all morphisms to an object in Im D and terminal
among all morphisms in (L). The following statement makes this precise.
Corollary 2.5.4. Let L : C → C be a localization functor and η : Id C → L a

morphism such that Lη is invertible. Then for each morphism ηX : X → LX
the following holds.
(1) The object LX belongs to Im L and every morphism X → Y with Y in

Im L factors uniquely through ηX.
(2) The morphism ηX belongs to (L) and factors uniquely through every
morphism X → Y in (L).
Proof. Apply Proposition 2.5.2.
Remark 2.5.5. (1) Let L : C → C be a localization functor and suppose there

are two morphisms ηi : Id C → L such that Lηi is invertible for i = 1, 2.
∼
Then there exists a unique isomorphism φ : L − → L such that η2 = φ ◦ η1 . This
follows from Lemma 2.5.1.
(2) Given any functor F : C → D, the full subcategory of F -local objects is
closed under taking all limits which exist in C.
2.6. Existence of localization functors

We provide a criterion for the existence of a localization functor L; it explains
how L is determined by the category of L-local objects.
Proposition 2.6.1. Let C be a category and D a full subcategory. Suppose that

every object in C isomorphic to one in D belongs to D. Then the following are
equivalent.
(1) There exists a localization functor L : C → C with Im L = D.

(2) For every object X in C there exists a morphism ηX : X → X with X in
D such that every morphism X → Y with Y in D factors uniquely through
ηX.
(3) The inclusion functor D → C admits a left adjoint.
Proof. (1) ⇒ (2): Suppose there exists a localization functor L : C → C with

Im L = D and let η : Id C → L be a morphism such that Lη is invertible. Then
Proposition 2.5.2 shows that C(ηX, Y ) is bijective for all Y in D.
(2) ⇒ (3): The morphisms ηX provide a functor F : C → D by sending
each X in C to X . It is straightforward to check that F is a left adjoint for the
inclusion D → C.
(3) ⇒ (1): Let G : D → C denote the inclusion and F its right adjoint. Then
L = G ◦ F is a localization functor with Im L = D by Proposition 2.4.1.
2.7. Localization functors preserving coproducts

We characterize the fact that a localization functor preserves small coproducts.
Proposition 2.7.1. Let L : C → C be a localization functor and suppose the

category C admits small coproducts. Then the following are equivalent.
170 Henning Krause
(1) The functor L preserves small coproducts.

(2) The L-local objects are closed under taking small coproducts in C.
(3) The right adjoint of the quotient functor C → C[(L)−1 ] preserves small
coproducts.
Proof. (1) ⇒ (2): Let (Xi )i∈I be a family of L-local objects. Thus the natural
morphisms Xi → LXi are invertible by Proposition 2.5.2 and they induce an
isomorphism
& ∼ & ∼ &
Xi −→ LXi −→ L( Xi ).
i i i
%
It follows that i Xi is L-local.
(2) ⇔ (3): We can identify C[(L)−1 ] = Im L by Corollary 2.5.3 and then
the right adjoint of the quotient functor identifies with the inclusion Im L → C.
Thus the right adjoint preserves small coproducts if and only if the inclusion
Im L → C preserves small coproducts.
(3) ⇒ (1): Write L as composite C − → C[(L)−1 ] − → C of the quotient
functor Q with its right adjoint L̄. Then Q preserves small coproducts since
it is a left adjoint. It follows that L preserves small coproducts if L̄ preserves
small coproducts.
2.8. Colocalization functors

A functor Γ : C → C is called colocalization functor if its opposite functor
Γ op : C op → C op is a localization functor. We call an object X in C Γ -colocal
if it is Γ op -local when viewed as an object of C op . Note that a colocalization
functor Γ : C → C induces an equivalence
∼
C[(Γ )−1 ] −→ Im Γ
and the essential image Im Γ equals the full subcategory of C consisting of all
Γ -colocal objects.
Remark 2.8.1. We think of Γ as L turned upside down; this explains our nota-
tion. Another reason for the use of Γ is the interpretation of local cohomology
as colocalization.
2.9. Example: Localization of modules

Let A be an associative ring and denote by Mod A the category of (right) A-
modules. Suppose that A is commutative and let S ⊆ A be a multiplicatively
closed subset, that is, 1 ∈ S and st ∈ S for all s, t ∈ S. We denote by

S −1 A = {x/s | x ∈ A and s ∈ S}
the ring of fractions. For each A-module M, let
S −1 M = {x/s | x ∈ M and s ∈ S}
be the localized module. An S −1 A-module N becomes an A-module via restric-
tion of scalars along the canonical ring homomorpism A → S −1 A. We obtain
a pair of functors
F : Mod A −→ Mod S −1 A, M
→ S −1 M ∼
= M ⊗A S −1 A,
G : Mod S −1 A −→ Mod A, N
→ N ∼
= HomS −1 A (S −1 A, N ).
Moreover, for each pair of modules M over A and N over S −1 A, we have
natural morphisms
ηM : M −→ (G ◦ F )M = S −1 M, x
→ x/1,
θ N : S −1 N = (F ◦ G)N −→ N, x/s
→ xs −1 .
These natural morphisms induce mutually inverse bijections as follows:
∼
HomA (M, GN) −→ HomS −1 A (F M, N ), α
→ θ N ◦ F α,
∼
HomS −1 A (F M, N ) −→ HomA (M, GN), β
→ Gβ ◦ ηM.
It is clear that the functors F and G form an adjoint pair, that is, F is
a left adjoint of G and G is a right adjoint of F . Moreover, the adjunction
morphism θ : F ◦ G → Id is invertible. Therefore the composite L = G ◦ F is
a localization functor.
Let us formulate this slightly more generally. Fix a ring homomorphism
f : A → B. Then it is well known that the restriction functor Mod B → Mod A
is fully faithful if and only if f is an epimorphism; see [45, Proposition XI.1.2].
Thus the functor Mod A → Mod A taking a module M to M ⊗A B is a local-
ization functor provided that f is an epimorphism.
2.10. Example: Localization of spectra

A spectrum E is a sequence of based topological spaces En and based home-
omorphisms En → #En+1 . A morphism of spectra E → F is a sequence of
based continuous maps En → Fn strictly compatible with the given struc-
tural homeomorphisms. The homotopy groups of a spectrum E are the groups
πn E = πn+i (Ei ) for i ≥ 0 and n + i ≥ 0. A morphism between spectra is a
172 Henning Krause
weak equivalence if it induces an isomorphism on homotopy groups. The stable

homotopy category Ho S is obtained from the category S of spectra by formally
inverting the weak equivalences. Thus Ho S = S[ −1 ] where denotes the
set of weak equivalences. We refer to [2, 39] for details.
2.11. Notes
The category of fractions is introduced by Gabriel and Zisman in [18], but the
idea of formally inverting elements can be traced back much further; see for
instance [36]. The appropriate context for localization functors is the theory of
monads; see [27].
3. Calculus of fractions
3.1. Calculus of fractions
Let C be a category and a set of morphisms in C. The category of fractions
C[ −1 ] admits an elementary description if some extra assumptions on are
satisfied. We say that admits a calculus of left fractions if the following
holds.
(LF1) If σ, τ are composable morphisms in , then τ ◦ σ is in . The identity

morphism id X is in for all X in C.
σ α
(LF2) Each pair of morphisms X ← −X− → Y with σ in can be completed
to a commutative square
X
α /Y
σ σ

X
α / Y
such that σ is in .
(LF3) Let α, β : X → Y be morphisms in C. If there is a morphism σ : X → X
in with α ◦ σ = β ◦ σ , then there exists a morphism τ : Y → Y in
with τ ◦ α = τ ◦ β.
Now assume that admits a calculus of left fractions. Then one obtains a
new category −1 C as follows. The objects are those of C. Given objects X
and Y , we call a pair (α, σ ) of morphisms
X
α / Y o σ
Y
in C with σ in a left fraction. The morphisms X → Y in −1 C are equivalence

classes [α, σ ] of such left fractions, where two diagrams (α1 , σ1 ) and (α2 , σ2 )
are equivalent if there exists a commutative diagram
? Y1 _??
?? σ1
α1
??
??
α3
X? / Y3 o σ3
Y
?? O
??
?
α2 ?? σ
2
Y2
with σ3 in . The composition of two equivalence classes [α, σ ] and [β, τ ] is
by definition the equivalene class [β ◦ α, σ ◦ τ ] where σ and β are obtained
from condition (LF2) as in the following commutative diagram.

Z
β }}
}> `BB
BB σ
} BB
}} BB
}}

> Y `BB Z `
~ BB σ β ||
|> @ @@
α ~~ BB @@τ
~~ || @@
~ BB ||
~ |
X Y Z
We obtain a canonical functor
P : C −→ −1 C
by taking the identity map on objects and by sending a morphism α : X → Y
to the equivalence class [α, id Y ]. Let us compare P with the quotient functor
Q : C → C[ −1 ].
Proposition 3.1.1. The functor F : −1 C → C[ −1 ] which is the identity
map on objects and which takes a morphism [α, σ ] to (Q σ )−1 ◦ Q α is an
isomorphism.
Proof. The functor P inverts all morphisms in and factors therefore through
Q via a functor G : C[ −1 ] → −1 C. It is straightforward to check that
F ◦ G = Id and G ◦ F = Id.
From now on, we will identify −1 C with C[ −1 ] whenever admits a
calculus of left fractions. A set of morphisms in C admits a calculus of right
fractions if the dual conditions of (LF1) – (LF3) are satisfied. Moreover,
is called a multiplicative system if it admits both, a calculus of left fractions
and a calculus of right fractions. Note that all results about sets of morphisms
174 Henning Krause
admitting a calculus of left fractions have a dual version for sets of morphisms
admitting a calculus of right fractions.
3.2. Calculus of fractions and adjoint functors

Given a category C and a set of morphisms , it is an interesting question to
ask when the quotient functor C → C[ −1 ] admits a right adjoint. It turns out
that this problem is closely related to the property of to admit a calculus of
left fractions.
Lemma 3.2.1. Let F : C → D and G : D → C be a pair of adjoint functors.

Assume that the right adjoint G is fully faithful and let be the set of morphisms
σ in C such that F σ is invertible. Then admits a calculus of left fractions.
Proof. We need to check the conditions (LF1) – (LF3). Observe first that
L = G ◦ F is a localization functor so that we can apply Proposition 2.5.2.
(LF1): This condition is clear because F is a functor.
σ α
(LF2): Let X ←−X− → Y be a pair of morphisms with σ in . This can be
completed to a commutative square
X
α /Y
σ σ
α

X / Y
if we take for σ the morphism ηY : Y → LY in , because the map C(σ, LY )

is surjective by Proposition 2.5.2.
(LF3): Let α, β : X → Y be morphisms in C and suppose there is a morphism
σ : X → X in with α ◦ σ = β ◦ σ . Then we take τ = ηY in and have
τ ◦ α = τ ◦ β, because the map C(σ, LY ) is injective by Proposition 2.5.2.
Lemma 3.2.2. Let C be a category and a set of morphisms admitting a

calculus of left fractions. Then the following are equivalent for an object X in
C.
(1) X is local with respect to .

(2) The quotient functor induces a bijection C(W, X) → C[ −1 ](W, X) for all
W.
Proof. (1) ⇒ (2): To show that fW : C(W, X) → C[ −1 ](W, X) is surjective,

α σ
→ X ←
choose a left fraction W − − X with σ in . Then there exists τ : X → X
with τ ◦ σ = id X since X is local. Thus fW (τ ◦ α) = [α, σ ]. To show that fW
is injective, suppose that fW (α) = fW (β). Then we have σ ◦ α = σ ◦ β for
some σ : X → X in . The morphism σ is a section because X is local, and

therefore α = β.
(2) ⇒ (1): Let σ : W → W be a morphism in . Then we have
C(σ, X) ∼= C[ −1 ]([σ, id W ], X). Thus C(σ, X) is bijective since [σ, id W ] is
invertible.
Proposition 3.2.3. Let C be a category, a set of morphisms admitting a

calculus of left fractions, and Q : C → C[ −1 ] the quotient functor. Then the
following are equivalent.
(1) The functor Q has a right adjoint (which is then fully faithful).
(2) For each object X in C, there exist a morphism ηX : X → X such that X
is local with respect to and Q(ηX) is invertible.
Proof. (1) ⇒ (2): Denote by Qρ the right adjoint of Q and by η : Id C →

Qρ Q the adjunction morphism. We take for each object X in C the morphism
ηX : X → Qρ QX. Note that Qρ QX is local by Proposition 2.5.2.
(2) ⇒ (1): We fix objects X and Y . Then we have two natural bijections
∼ ∼
C[ −1 ](X, Y ) −
→ C[ −1 ](X, Y ) ←
− C(X, Y ).
The first is induced by ηY : Y → Y and is bijective since Q(ηY ) is invertible.

The second map is bijective by Lemma 3.2.2, since Y is local with respect to
. Thus we obtain a right adjoint for Q by sending each object Y of C[ −1 ]
to Y .
3.3. A criterion for the fractions to form a small set

Let C be a category and a set of morphisms in C. Suppose that admits a
calculus of left fractions. From the construction of C[ −1 ] we cannot expect
that for any given pair of objects X and Y the equivalence classes of fractions
in C[ −1 ](X, Y ) form a small set. The situation is different if the category C
is small. Then it is clear that C[ −1 ](X, Y ) is a small set for all objects X, Y .
The following criterion generalizes this simple observation.
Lemma 3.3.1. Let C be a category and a set of morphisms in C which

admits a calculus of left fractions. Let Y be an object in C and suppose that
there exists a small set S = S(Y, ) of objects in C such that for every morphism
σ : Y → Y in there is a morphism τ : Y → Y with τ ◦ σ in and Y in
S. Then C[ −1 ](X, Y ) is a small set for every object X in C.
176 Henning Krause
α σ
Proof. The condition on Y implies that every fraction X → Y ← Y is equiv-
α σ
alent to one of the form X → Y ← Y with Y in S. Clearly, the fractions of
the form (α , σ ) with σ ∈ C(Y, Y ) and Y ∈ S form a small set.
3.4. Calculus of fractions for subcategories

We provide a criterion such that the calculus of fractions for a set of morphisms
in a category C is compatible with the passage to a subcategory of C.
Lemma 3.4.1. Let C be a category and a set of morphisms admitting

a calculus of left fractions. Suppose D is a full subcategory of C such
that for every morphism σ : Y → Y in with Y in D there is a mor-
phism τ : Y → Y with τ ◦ σ in ∩ D. Then ∩ D admits a calculus
of left fractions and the induced functor D[( ∩ D)−1 ] → C[ −1 ] is fully
faithful.
Proof. It is straightforward to check (LF1) – (LF3) for ∩ D. Now let X, Y

be objects in D. Then we need to show that the induced map
f : D[( ∩ D)−1 ](X, Y ) −→ C[ −1 ](X, Y )
is bijective. The map sends the equivalence class of a fraction to the equivalence
class of the same fraction. If [α, σ ] belongs to C[ −1 ](X, Y ) and τ is a mor-
phism with τ ◦ σ in ∩ D, then [τ ◦ α, τ ◦ σ ] belongs to D[( ∩ D)−1 ](X, Y )
and f sends it to [α, σ ]. Thus f is surjective. A similar argument shows that f
is injective.
Example 3.4.2. Let A be a commutative noetherian ring and S ⊆ A a mul-

tiplicatively closed subset. Denote by the set of morphisms σ in Mod A
such that S −1 σ is invertible. Then is a multiplicative system and one can
show directly that for the subcategory mod A of finitely generated A-modules
and T = ∩ mod A the dual of the condition in Lemma 3.4.1 holds. Thus the
induced functor
(mod A)[T −1 ] −→ (Mod A)[ −1 ]
is fully faithful.
3.5. Calculus of fractions and coproducts

We provide a criterion for the quotient functor C → C[ −1 ] to preserve small
coproducts.
Proposition 3.5.1. Let C be a category which admits small coproducts. Suppose

that is a set of morphisms in C which admits a calculus of left fractions. If
% −1
i σi belongs to for every family (σi )i∈I in , then the category C[ ]
−1
admits small coproducts and the quotient functor C → C[ ] preserves small
coproducts.
Proof. Let (Xi )i∈I be a family of objects in C[ −1 ] which is indexed by a small
%
set I . We claim that the coproduct i Xi in C is also a coproduct in C[ −1 ].
Thus we need to show that for every object Y , the canonical map
&
C[ −1 ]( Xi , Y ) −→ C[ −1 ](Xi , Y ) (3.5.1)
i i
is bijective.
αi σi
To check surjectivity of (3.5.1), let (Xi → Zi ← Y )i∈I be a family of left
fractions. Using (LF2), we obtain a commutative diagram
% %
% αi % σi %
i
/ o i
i Xi i Zi i Y
πY

Zo
σ
Y
%
where πY : i Y → Y is the summation morphism and σ ∈ . It is easily
checked that
σ αi σi
(Xi → Z ← Y ) ∼ (Xi → Zi ← Y )
% σ αi
for all i ∈ I , and therefore (3.5.1) sends i Xi → Z ← Y to the family (Xi →
σi
Zi ← Yi )i∈I .
% α σ % α σ
To check injectivity of (3.5.1), let i Xi → Z ← Y and i Xi → Z ← Y
be left fraction such that
αi σ αi σ
(Xi → Z ← Y ) ∼ (Xi → Z ← Y )
for all i. We may assume that Z = Z = Z and σ = σ = σ since we can

choose morphisms τ : Z → Z and τ : Z → Z with τ ◦ σ = τ ◦ σ ∈ .
Thus there are morphisms βi : Z → Zi with βi ◦ αi = βi ◦ αi and βi ◦ σ ∈
for all i. Each βi belongs to the saturation ¯ of which is the set of all
morphisms in C which become invertible in C[ −1 ]. Note that a morphism φ
¯ if and only if there are morphisms φ and φ such that φ ◦ φ
in C belongs to

and φ ◦ φ belong to . Therefore ¯ is also closed under taking coproducts.
Moreover, ¯ admits a calculus of left fractions, and we obtain therefore a
178 Henning Krause
commutative diagram
% %
Xi / Z
πZ
/Z
i i
%
i βi τ
%
Zi / Z∗
i
with τ ∈ .
¯ Thus τ ◦ σ ∈ ,¯ and we have
& α σ & α σ
( Xi → Z ← Y ) ∼ ( Xi → Z ← Y )
i i
% %
since πZ ◦ i αi = α and πZ ◦ i αi = α . Therefore the map (3.5.1) is also
injective, and this completes the proof.
Example 3.5.2. Let C be a category which admits small coproducts and

L : C → C be a localization functor. Then a morphism σ in C belongs to
= (L) if and only if the induced map C(σ, LX) is invertible for every
object X in C. Thus is closed under taking small coproducts and therefore
the quotient functor C → C[ −1 ] preserves small coproducts.
3.6. Notes
The calculus of fractions for categories has been developed by Gabriel and
Zisman in [18] as a tool for homotopy theory.
4. Localization for triangulated categories

4.1. Triangulated categories
Let T be an additive category with an equivalence S : T → T . A triangle in
T is a sequence (α, β, γ ) of morphisms
α β γ
X −→ Y −→ Z −→ SX,
and a morphism between two triangles (α, β, γ ) and (α , β , γ ) is a triple

(φ1 , φ2 , φ3 ) of morphisms in T making the following diagram commutative.
X
α /Y β
/Z γ
/ SX
φ1 φ2 φ3 Sφ1

X
α / Y β
/ Z γ
/ SX
The category T is called triangulated if it is equipped with a set of distinguished

triangles (called exact triangles) satisfying the following conditions.
(TR1) A triangle isomorphic to an exact triangle is exact. For each object X,

id
the triangle 0 → X − → X → 0 is exact. Each morphism α fits into an
exact triangle (α, β, γ ).
(TR2) A triangle (α, β, γ ) is exact if and only if (β, γ , −Sα) is exact.
(TR3) Given two exact triangles (α, β, γ ) and (α , β , γ ), each pair of mor-
phisms φ1 and φ2 satisfying φ2 ◦ α = α ◦ φ1 can be completed to a
morphism
X
α /Y β
/Z γ
/ SX
φ1 φ2 φ3 Sφ1
α
β γ
X / Y / Z / SX
of triangles.
(TR4) Given exact triangles (α1 , α2 , α3 ), (β1 , β2 , β3 ), and (γ1 , γ2 , γ3 ) with
γ1 = β1 ◦ α1 , there exists an exact triangle (δ1 , δ2 , δ3 ) making the fol-
lowing diagram commutative.
X
α1
/Y α2
/U α3
/ SX
β1 δ1

X
γ1
/Z γ2
/V γ3
/ SX
β2 δ2 Sα1

W W
β3
/ SY
β3 δ3

SY
Sα2
/ SU
Recall that an idempotent endomorphism φ = φ 2 of an object X in an

π ι
additive category splits if there exists a factorization X −
→Y −
→ X of φ with
π ◦ ι = id Y .
Remark 4.1.1. Suppose a triangulated category T admits countable coprod-

ucts. Then every idempotent endomorphism splits. More precisely, let φ : X →
X be an idempotent morphism in T , and denote by Y a homotopy colimit of
the sequence
φ φ φ
X −→ X −→ X −→ · · · .
180 Henning Krause
The morphism φ factors through the canonical morphism π : X → Y via a

morphism ι : Y → X, and we have π ◦ ι = id Y . Thus φ splits; see [33, Propo-
sition 1.6.8] for details.
4.2. Exact functors

An exact functor T → U between triangulated categories is a pair (F, μ)
consisting of a functor F : T → U and an isomorphism μ : F ◦ ST → SU ◦ F
α β γ
such that for every exact triangle X → Y → Z → ST X in T the triangle
Fα Fβ μX ◦ F γ
F X −→ F Y −→ F Z −−−−→ SU (F X)
is exact in U.
We have the following useful lemma.
Lemma 4.2.1. Let F : T → U and G : U → T be an adjoint pair of functors

between triangulated categories. If one of both functors is exact, then also the
other is exact.
Proof. See [33, Lemma 5.3.6].
4.3. Multiplicative systems

Let T be a triangulated category and a set of morphisms which is a multi-
plicative system. Recall this means that admits a calculus of left and right
fractions. Then we say that is compatible with the triangulation if
(1) given σ in , the morphism S n σ belongs to for all n ∈ Z, and

(2) given a morphism (φ1 , φ2 , φ3 ) between exact triangles with φ1 and φ2 in
, there is also a morphism (φ1 , φ2 , φ3 ) with φ3 in .
Lemma 4.3.1. Let T be a triangulated category and a multiplicative system

of morphisms which is compatible with the triangulation. Then the quotient
category T [ −1 ] carries a unique triangulated structure such that the quotient
functor T → T [ −1 ] is exact.
Proof. The equivalence S : T → T induces a unique equivalence T [ −1 ] →

T [ −1 ] which commutes with the quotient functor Q : T → T [ −1 ]. This
follows from the fact that S = . Now take as exact triangles in T [ −1 ]
all those isomorphic to images of exact triangles in T . It is straightforward to
verify the axioms (TR1) – (TR4); see [48, II.2.2.6]. The functor Q is exact by
construction. In particular, we have Q ◦ ST = ST [ −1 ] ◦ Q.
4.4. Cohomological functors

A functor H : T → A from a triangulated category T to an abelian category
A is cohomological if H sends every exact triangle in T to an exact sequence
in A.
Example 4.4.1. For each object X in T , the representable functors

T (X, −) : T → Ab and T (−, X) : T op → Ab into the category Ab of abelian
groups are cohomological functors.
Lemma 4.4.2. Let H : T → A be a cohomological functor. Then the set

of morphisms σ in T such that H (S n σ ) is invertible for all n ∈ Z forms a
multiplicative system which is compatible with the triangulation of T .
Proof. We need to verify that admits a calculus of left and right fractions.
In fact, it is sufficient to check conditions (LF1) – (LF3), because then the dual
conditions are satisfied as well since the definition of is self-dual.
(LF1): This condition is clear because H is a functor.
(LF2): Let α : X → Y and σ : X → X be morphisms with σ in . We com-
plete α to an exact triangle and apply (TR3) to obtain the following morphism
between exact triangles.
W /X α /Y / SW
σ σ

W / X α / Y / SW
Then the 5-lemma shows that σ belongs to .

(LF3): Let α, β : X → Y be morphisms in T and σ : X → X in such that
σ φ
α ◦ σ = β ◦ σ . Complete σ to an exact triangle X → X → X → SX . Then
α − β factors through φ via some morphism ψ : X → Y . Now complete
ψ τ
ψ to an exact triangle X → Y → Y → SX . Then τ belongs to and
τ ◦ α = τ ◦ β.
It remains to check that is compatible with the triangulation. Condition
(1) is clear from the definition of . For condition (2), observe that given any
morphism (φ1 , φ2 , φ3 ) between exact triangles with φ1 and φ2 in , we have
that φ3 belongs to . This is an immediate consequence of the 5-lemma.
4.5. Triangulated and thick subcategories

Let T be a triangulated category. A non-empty full subcategory S is a triangu-
lated subcategory if the following conditions hold.
182 Henning Krause
(TS1) S n X ∈ S for all X ∈ S and n ∈ Z.

(TS2) Let X → Y → Z → SX be an exact triangle in T . If two objects from
{X, Y, Z} belong to S, then also the third.
A triangulated subcategory S is thick if in addition the following condition

holds.
π ι
(TS3) Let X −
→Y − → X be morphisms in T such that id Y = π ◦ ι. If X
belongs to S, then also Y .
Note that a triangulated subcategory S of T inherits a canonical triangulated

structure from T .
Next observe that a triangulated subcategory S of T is thick provided that S
admits countable coproducts. This follows from the fact that in a triangulated
category with countable coproducts all idempotent endomorphisms split.
Let T be a triangulated category and let F : T → U be an additive functor.
The kernel Ker F of F is by definition the full subcategory of T which is formed
by all objects X such that F X = 0. If F is an exact functor into a triangulated
category, then Ker F is a thick subcategory of T . Also, if F is a cohomological
'
functor into an abelian category, then n∈Z S n (Ker F ) is a thick subcategory
of T .
4.6. Verdier localization

Let T be a triangulated category. Given a triangulated subcategory S, we denote
by (S) the set of morphisms X → Y in T which fit into an exact triangle
X → Y → Z → SX with Z in S.
Lemma 4.6.1. Let T be a triangulated category and S a triangulated sub-

category. Then (S) is a multiplicative system which is compatible with the
triangulation of T .
Proof. The proof is similar to that of Lemma 4.4.2; see [48, II.2.1.8] for
details.
The localization of T with respect to a triangulated subcategory S is by

definition the quotient category
T /S := T [(S)−1 ]
together with the quotient functor T → T /S.

Proposition 4.6.2. Let T be a triangulated category and S a full triangulated

subcategory. Then the category T /S and the quotient functor Q : T → T /S
have the following properties.
(1) The category T /S carries a unique triangulated structure such that Q is

exact.
(2) A morphism in T is annihilated by Q if and only if it factors through an
object in S.
(3) The kernel Ker Q is the smallest thick subcategory containing S.
(4) Every exact functor T → U annihilating S factors uniquely through Q via
an exact functor T /S → U.
(5) Every cohomological functor T → A annihilating S factors uniquely
through Q via a cohomological functor T /S → A.
Proof. (1) follows from Lemma 4.3.1.

(2) Let φ be a morphism in T . We have Qφ = 0 iff σ ◦ φ = 0 for some
σ ∈ (S) iff φ factors through some object in S.
(3) Let X be an object in T . Then QX = 0 if and only if Q(id X) = 0. Thus
part (2) implies that the kernel of Q conists of all direct factors of objects in S.
(4) An exact functor F : T → U annihilating S inverts every morphism in
(S). Thus there exists a unique functor F̄ : T /S → U such that F = F̄ ◦ Q.
The functor F̄ is exact because an exact triangle in T /S is up to isomorphism
of the form Q for some exact triangle in T . Thus F̄ ∼ = F is exact.
(5) Analogous to (4).
4.7. Localization of subcategories

Let T be a triangulated category with two full triangulated subcategories T
and S. Then we put S = S ∩ T and have T (S ) = T (S) ∩ T . Thus we
can form the following commutative diagram of exact functors
S
inc /T can / T /S
inc inc J

S
inc /T can / T /S
and ask when the functor J is fully faithful. We have the following criterion.
Lemma 4.7.1. Let T , T , S, S be as above. Suppose that either
(1) every morphism from an object in S to an object in T factors through some

object in S , or
184 Henning Krause
(2) every morphism from an object in T to an object in S factors through some

object in S .
Then the induced functor J : T /S → T /S is fully faithful.
Proof. Suppose that condition (1) holds. We apply the criterion from
Lemma 3.4.1. Thus we take a morphism σ : Y → Y from (S) with Y in
T and need to find τ : Y → Y such that τ ◦ σ belongs to (S) ∩ T . To this
φ σ
end complete σ to an exact triangle X − → Y → SX. Then X belongs
→Y −
φ φ
to S and by our assumption we have a factorization X −
→Z−
→ Y of φ with
φ ψ
Z in S . Complete φ to an exact triangle Z −
→Y − → Y → SZ. Then (TR3)
yields a morphism τ : Y → Y satisfying ψ = τ ◦ σ . In particular, τ ◦ σ lies
in (S) ∩ T since Z belongs to S . The proof using condition (2) is dual.
4.8. Orthogonal subcategories

Let T be a triangulated category and S a triangulated subcategory. Then we
define two full subcategories
S ⊥ = {Y ∈ T | T (X, Y ) = 0 for all X ∈ S}

⊥
S = {X ∈ T | T (X, Y ) = 0 for all Y ∈ S}
and call them orthogonal subcategories with respect to S. Note that S ⊥ and
⊥
S are thick subcategories of T .
Lemma 4.8.1. Let T be a triangulated category and S a triangulated subcat-

egory. Then the following are equivalent for an object Y in T .
(1) Y belongs to S ⊥ .
(2) Y is (S)-local, that is, T (σ, Y ) is bijective for all σ in (S).
(3) The quotient functor induces a bijection T (X, Y ) → T /S(X, Y ) for all X
in T .
Proof. (1) ⇒ (2): Suppose T (X, Y ) = 0 for all X in S. Then every σ in (S)
induces a bijection T (σ, Y ) because T (−, Y ) is cohomological. Thus Y is
(S)-local.
(2) ⇒ (1): Suppose that Y is (S)-local. If X belongs to S, then the mor-
phism σ : X → 0 belongs to (S) and induces therefore a bijection C(σ, Y ).
Thus Y belongs to S ⊥ .
(2) ⇔ (3): Apply Lemma 3.2.2.
4.9. Bousfield localization

Let T be a triangulated category. We wish to study exact localization functors
L : T → T . To be more precise, we assume that L is an exact functor and that L
is a localization functor in the sense that there exists a morphism η : Id C → L
with Lη : L → L2 being invertible and Lη = ηL. Note that there is an isomor-
∼
phism μ : L ◦ S − → S ◦ L since L is exact, and there exists a unique choice such
that μX ◦ ηSX = SηX for all X in T . This follows from Lemma 2.5.1.
We observe that the kernel of an exact localization functor is a thick subcat-
egory of T . The following fundamental result characterizes the thick subcate-
gories of T which are of this form.
Proposition 4.9.1. Let T be a triangulated category and S a thick subcategory.

Then the following are equivalent.
(1) There exists an exact localization functor L : T → T with Ker L = S.

(2) The inclusion functor S → T admits a right adjoint.
(3) For each X in T there exists an exact triangle X → X → X → SX
with X in S and X in S ⊥ .
(4) The quotient functor T → T /S admits a right adjoint.
inc can
(5) The composite S ⊥ −→ T −→ T /S is an equivalence.
(6) The inclusion functor S → T admits a left adjoint and ⊥ (S ⊥ ) = S.
⊥
Proof. Let I : S → T and J : S ⊥ → T denote the inclusions and Q : T →

T /S the quotient functor.
(1) ⇒ (2): Suppose that L : T → T is an exact localization functor with
Ker L = S and let η : Id T → L be a morphism such that Lη is invertible. We
obtain a right adjoint Iρ : T → S for the inclusion I by completing for each
θX ηX
X in T the morphism ηX to an exact triangle Iρ X −→ X −→ LX → S(Iρ X).
Note that Iρ X belongs to S since LηX is invertible. Moreover, T (W, θ X)
is bijective for all W in S since T (W, LX) = 0 by Lemma 4.8.1. Here we
use that LX is (L)-local by Proposition 2.5.2 and that (L) = (S). Thus
Iρ provides a right adjoint for I since T (W, Iρ X) ∼
= T (I W, X) for all W in
S and X in T . In particular, we see that the exact triangle defining Iρ X is,
up to a unique isomorphism, uniquely determined by X. Therefore Iρ is well
defined.
(2) ⇒ (3): Suppose that Iρ : T → S is a right adjoint of the inclusion I . We
fix an object X in T and complete the adjunction morphism θ X : Iρ X → X
θX
to an exact triangle Iρ X −→ X → X → S(Iρ X). Clearly, Iρ X belongs to S.
We have T (W, X ) = 0 for all W in S since T (W, θ X) is bijective. Thus X
belongs to S ⊥ .
186 Henning Krause
(3) ⇒ (4): We apply Proposition 3.2.3 to obtain a right adjoint for

the quotient functor Q. To this end fix an object X in T and an exact
η
triangle X → X − → X → SX with X in S and X in S ⊥ . The mor-
phism η belongs to (S) by definition, and the object X is (S)-local by
Lemma 4.8.1. Now it follows from Proposition 3.2.3 that Q admits a right
adjoint.
(4) ⇒ (1): Let Qρ : T /S → T denote a right adjoint of Q. This functor is
fully faithful by Proposition 2.3.1 and exact by Lemma 4.2.1. Thus L = Qρ ◦ Q
is an exact functor with Ker L = Ker Q = S. Moreover, L is a localization
functor by Corollary 2.4.2.
(4) ⇒ (5): Let Qρ : T /S → T denote a right adjoint of Q. The composite
Q ◦ J : S ⊥ → T /S is fully faithful by Lemma 4.8.1. Given an object X in
T /S, we have Q(Qρ X) ∼ = X by Proposition 2.3.1, and Qρ X belongs to S ⊥ ,
since T (W, Qρ X) ∼ = T /S(QW, X) = 0 for all W in S. Thus Q ◦ J is dense
and therefore an equivalence.
(5) ⇒ (6): Suppose Q ◦ J : S ⊥ → T /S is an equivalence and let
F : T /S → S ⊥ be a quasi-inverse. We have for all X in T and Y in S ⊥
∼ ∼ ∼
T (X, J Y ) − → S ⊥ (F QX, F QJ Y ) −
→ T /S(QX, QJ Y ) − → S ⊥ (F QX, Y ),
where the first bijection follows from Lemma 4.8.1 and the others are clear
from the choice of F . Thus F ◦ Q is a left adjoint for the inclusion J .
It remains to show that ⊥ (S ⊥ ) = S. The inclusion ⊥ (S ⊥ ) ⊇ S is clear. Now
let X be an object of ⊥ (S ⊥ ). Then we have
T /S(QX, QX) ∼
= S ⊥ (F QX, F QX) ∼
= T (X, J (F QX)) = 0.
Thus QX = 0 and therefore X belongs to S.

(6) ⇒ (3): Suppose that Jλ : T → S ⊥ is a left adjoint of the inclusion J . We
fix an object X in T and complete the adjunction morphism μX : X → Jλ X
μX
to an exact triangle X → X −→ Jλ X → SX . Clearly, Jλ X belongs to S ⊥ .
We have T (X , Y ) = 0 for all Y in S ⊥ since T (μX, Y ) is bijective. Thus X
belongs to ⊥ (S ⊥ ) = S.
The following diagram displays the functors which arise from a localization
functor L : T → T . We use the convention that Fρ denotes a right adjoint of a
functor F .
Iρ Qρ
o o
S / T / T /S (L = Qρ ◦ Q and Γ = I ◦ Iρ )
I =inc Q=can
4.10. Acyclic and local objects

Let T be a triangulated category and L : T → T an exact localization functor.
An object X in T is by definition L-acyclic if LX = 0. Recall that an object
in T is L-local if and only if it belongs to the essential image Im L of L;
see Proposition 2.5.2. The exactness of L implies that S := Ker L is a thick
subcategory and that (L) = (S). Therefore L-local and (S)-local objects
coincide.
The following result says that acyclic and local objects form an orthogonal
pair.
Proposition 4.10.1. Let L : T → T be an exact localization functor. Then we

have
Ker L = ⊥ (Im L) and (Ker L)⊥ = Im L.
More explictly, the following holds.
(1) X ∈ T is L-acyclic if and only if T (X, Y ) = 0 for every L-local object Y .

(2) Y ∈ T is L-local if and only if T (X, Y ) = 0 for every L-acyclic object X.
Proof. (1) We write L = G ◦ F where F is a functor and G a fully faithful right

adjoint; see Corollary 2.4.2. Suppose first we have given objects X, Y such that
X is L-acyclic and Y is L-local. Observe that F X = 0 since G is faithful. Thus
T (X, Y ) ∼
= T (X, GF Y ) ∼
= T (F X, F Y ) = 0.
Now suppose that X is an object with T (X, Y ) = 0 for all L-local Y . Then
T (F X, F X) ∼
= T (X, GF X) = 0
and therefore F X = 0. Thus X is L-acyclic.

(2) This is a reformulation of Lemma 4.8.1.
4.11. A functorial triangle

Let T be a triangulated category and L : T → T an exact localization functor.
We denote by η : Id T → L a morphism such that Lη is invertible. It follows
from Proposition 4.9.1 and its proof that we obtain an exact functor Γ : T → T
by completing for each X in T the morphism ηX to an exact triangle
θX ηX
Γ X −→ X −→ LX −→ S(Γ X). (4.11.1)
188 Henning Krause
The exactness of Γ follows from Lemma 4.2.1. Observe that Γ X is L-acyclic

and that LX is L-local. In fact, the exact triangle (4.11.1) is essentially deter-
mined by these properties. This is a consequence of the following basic prop-
erties of L and Γ .
Proposition 4.11.1. The functors L, Γ : T → T have the following properties.

∼
(1) L induces an equivalence T / Ker L − → Im L.
(2) L induces a left adjoint for the inclusion Im L → T .
(3) Γ induces a right adjoint for the inclusion Ker L → T .
Proof. (1) is a reformulation of Corollary 2.5.3, and (2) follows from Corol-
lary 2.4.2. (3) is an immediate consequence of the construction of Γ via Propo-
sition 4.9.1.
Proposition 4.11.2. Let L : T → T be an exact localization functor and X an

object in C. Given any exact triangle X → X → X → SX with X L-acyclic
and X L-local, there are unique isomorphisms α and β making the following
diagram commutative.
X /X / X / SX (4.11.2)

α β Sα

ΓX
θX /X ηX
/ LX / S(Γ X)
Proof. The morphism θ X induces a bijection T (X , θX) since X is acyclic.

Thus X → X factors uniquely through θ X via a morphism α : X → Γ X.
An application of (TR3) gives a morphism β : X → LX making the diagram
(4.11.2) commutative. Now apply L to this diagram. Then Lβ is an isomorphism
since LX = 0 = LΓ X, and Lβ is isomorphic to β since X and LX are L-
local. Thus β is an isomorphism, and therefore α is an isomorphism.
4.12. Localization versus colocalization

For exact functors on triangulated categories, we have the following symmetry
principle relating localization and colocalization.
Proposition 4.12.1. Let T be a triangulated category.
(1) Suppose L : T → T is an exact localization functor and Γ : T → T the

functor which is defined in terms of the exact triangle (4.11.1). Then Γ is
an exact colocalization functor with Ker Γ = Im L and Im Γ = Ker L.
(2) Suppose Γ : T → T is an exact colocalization functor and L : T → T

the functor which is defined in terms of the exact triangle (4.11.1). Then L
is an exact localization functor with Ker L = Im Γ and Im L = Ker Γ .
Proof. It suffices to prove (1) because (2) is the dual statement. So let L : T →
T be an exact localization functor. It follows from the construction of Γ that it
is of the form Γ = I ◦ Iρ where Iρ denotes a right adjoint of the fully faithful
inclusion I : Ker L → T . Thus Γ is a colocalization functor by Corollary 2.4.2.
The exactness of Γ follows from Lemma 4.2.1, and the identities Ker Γ = Im L
and Im Γ = Ker L are easily derived from the exact triangle (4.11.1).
4.13. Recollements
A recollement is by definition a diagram of exact functors
Iρ Qρ
o o
T o / T / T (4.13.1)
I o Q
Iλ Qλ
satisfying the following conditions.

(1) Iλ is a left adjoint and Iρ a right adjoint of I .
(2) Qλ is a left adjoint and Qρ a right adjoint of Q.
(3) Iλ I ∼
= Id T ∼= Iρ I and QQρ ∼ = Id T ∼
= QQλ .
(4) Im I = Ker Q.
Note that the isomorphisms in (3) are supposed to be the adjunction morphisms
resulting from (1) and (2).
A recollement gives rise to various localization and colocalization functors
for T . First observe that the functors I , Qλ , and Qρ are fully faithful; see
Proposition 2.3.1. Therefore Qρ Q and I Iλ are localization functors and Qλ Q
and I Iρ are colocalization functors. This follows from Corollary 2.4.2. Note that
the localization functor L = Qρ Q has the additional property that the inclusion
Ker L → T admits a left adjoint. Moreover, L determines the recollement up
to an equivalence.
Proposition 4.13.1. Let L : T → T be an exact localization functor and sup-

pose that the inclusion Ker L → T admits a left adjoint. Then L induces a
recollement of the following form.
o o
Ker L o / T / Im L
inc
o
Moreover, any recollement for T is, up to equivalences, of this form for some
exact localization functor L : T → T .
190 Henning Krause
Proof. We apply Proposition 4.9.1 and its dual assertion. Observe first that any
localization functor L : T → T induces the following diagram.
Iρ =Γ Qρ =inc
o o
Ker L / T / Im L
I =inc Q=L
The functor I admits a left adjoint if and only if Q admits a left adjoint. Thus
the diagram can be completed to a recollement if and only if the inclusion I
admits a left adjoint.
Suppose now there is given a recollement of the form (4.13.1). Then L =
Qρ Q is a localization functor and the inclusion Ker L → T admits a left
∼
adjoint. The functor I induces an equivalence T − → Ker L and Qρ induces
∼
an equivalence T −→ Im L. It is straightforward to formulate and check the
various compatibilities of these equivalences.
As a final remark, let us mention that for any recollement of the form (4.13.1),
the functors Qλ and Qρ provide two (in general different) embeddings of T
into T . If we identify T = Im I , then Qρ identifies T with (T )⊥ and Qλ
identifies T with ⊥ (T ); see Proposition 4.10.1.
4.14. Example: The derived category of a module category

Let A be an associative ring. We denote by K(Mod A) the category of chain
complexes of A-modules whose morphisms are the homotopy classes of chain
maps. The functor H n : K(Mod A) → Mod A taking the cohomology of a com-
plex in degree n is cohomological. A morphism φ is called quasi-isomorphism
if H n φ is an isomorphism for all n ∈ Z, and we denote the set of all quasi-
isomorphisms by qis. Then
D(A) := D(Mod A) := K(Mod A)[qis−1 ]
is by definition the derived category of Mod A. The kernel of the quotient

functor Q : K(Mod A) → D(Mod) is the full subcategory Kac (Mod A) which
is formed by all acyclic complexes. Note that Q admits a left adjoint Qλ
taking each complex to its K-projective resolution and a right adjoint Qρ
taking each complex to its K-injective resolution. Thus we obtain the following
recollement.
Qρ
o o
Kac (Mod A) o / K(Mod A) / D(Mod A) (4.14.1)
inc
o Q
Qλ
It follows that for each pair of chain complexes X, Y the set of mor-
phisms D(Mod A)(X, Y ) is small, since Qλ induces a bijection with
K(Mod A)(Qλ X, Qλ Y ). The adjoints of Q are discussed in more detail in

Section 5.8.
4.15. Example: A derived category without small morphism sets

For any abelian category A, the derived category D(A) is by definition
K(A)[qis−1 ]. Here, K(A) denotes the category of chain complexes in A whose
morphisms are the homotopy classes of chain maps, and qis denotes the set
of quasi-isomorphisms. Let us identify objects in A with chain complexes
concentrated in degree zero.
We give an example of an abelian category A and an object X in A such that
the set Ext1A (X, X) ∼
= D(A)(X, SX) is not small. This example is taken from
Freyd [14, pp. 131] and has been pointed out to me by Neeman.
Let U denote the set of all cardinals of small sets. This set is not small.
Consider the free associative Z-algebra ZU which is generated by the ele-
ments of U . Now let A = Mod A denote the category of A-modules, where it
is assumed that the underlying set of each module is small. Let Z denote the
trivial A-module, that is, zu = 0 for all z ∈ Z and u ∈ U . We claim that the
set Ext1A (Z, Z) is not small. To see this, define for each u ∈ U an A-module
Eu = Z ⊕ Z by

(z2 , 0) if x = u,
(z1 , z2 )x =
(0, 0) if x = u,
where
( )
(z1 , z2 ) ∈ Eu and x ∈ U . Then we have short exact sequences 0 →
1
0 [0 1]
Z −−→ Eu −−−→ Z → 0 which yield pairwise different elements of Ext1A (Z, Z)
as u runs though the elements in U .
4.16. Example: The recollement induced by an idempotent

Recollements can be defined for abelian categories in the same way as for
triangulated categories. A typical example arises for any module category from
an idempotent element of the underlying ring.
Let A be an associative ring and e2 = e ∈ A an idempotent. Then the func-
tor F : Mod A → Mod eAe taking a module M to Me and restriction along
p : A → A/AeA induce the following recollement.
o o HomeAe (Ae,−)
Mod A/AeA o p∗ / Mod A / Mod eAe
o F
−⊗eAe eA
192 Henning Krause
Note that we can describe adjoints of F since

F = HomA (eA, −) = − ⊗A Ae.
The recollement for Mod A induces the following recollement of triangulated
categories for D(A).
o o RHomeAe (Ae,−)
Ker D(F ) o / D(A) / D(eAe)
inc
o D(F )
−⊗LeAe eA
The functor F is exact and D(F ) takes by definition a complex X to F X.

The functor D(p∗ ) : D(A/AeA) → D(A) identifies D(A/AeA) with Ker D(F )
i (A/AeA, A/AeA) = 0 for all i > 0.
if and only if TorA
4.17. Notes
Triangulated categories were introduced independently in algebraic geome-
try by Verdier in his thèse [48], and in algebraic topology by Puppe [38].
Grothendieck and his school used the formalism of triangulated and derived
categories for studying homological properties of abelian categories. Early
examples are Grothendieck duality and local cohomology for categories of
sheaves. The basic example of a triangulated category from topology is the
stable homotopy category.
Localizations of triangulated categories are discussed in Verdier’s thèse
[48]. In particular, he introduced the localization (or Verdier quotient) of a
triangulated category with respect to a triangulated subcategory. In the context
of stable homotopy theory, it is more common to think of localization functors
as endofunctors; see for instance the work of Bousfield [8], which explains
the term Bousfield localization. The standard reference for recollements is [6].
Resolutions of unbounded complexes were first studied by Spaltenstein in [44];
see also [5].
5. Localization via Brown representability

5.1. Brown representatbility
Let T be a triangulated category and suppose that T has small coproducts. A
localizing subcategory of T is by definition a thick subcategory which is closed
under taking small coproducts. A localizing subcategory of T is generated by
a fixed set of objects if it is the smallest localizing subcategory of T which
contains this set.
We say that T is perfectly generated by some small set S of objects of T

provided the following holds.
(PG1) There is no proper localizing subcategory of T which contains S.

(PG2) Given a family (Xi → Yi )i∈I of morphisms in T such that the induced
map T (C, Xi ) → T (C, Yi ) is surjective for all C ∈ S and i ∈ I , the
induced map
& &
T (C, Xi ) −→ T (C, Yi )
i i
is surjective.
We say that a triangulated category T with small products is perfectly cogen-

erated if T op is perfectly generated.
Theorem 5.1.1. Let T be a triangulated category with small coproducts and

suppose T is perfectly generated.
(1) A functor F : T op → Ab is cohomological and sends small coproducts in

T to products if and only if F ∼= T (−, X) for some object X in T .
(2) An exact functor T → U between triangulated categories preserves small
coproducts if and only if it has a right adjoint.
Proof. For a proof of (1) see [24, Theorem A]. To prove (2), suppose that F
preserves small coproducts. Then one defines the right adjoint G : U → T by
sending an object X in U to the object in T representing U(F −, X). Thus
U(F −, X) ∼= T (−, GX). Conversely, given a right adjoint of F , it is automatic
that F preserves small coproducts.
Remark 5.1.2. (1) In the presence of (PG2), condition (PG1) is equivalent to

the following: For an object X in T , we have X = 0 if T (S n C, X) = 0 for all
C ∈ S and n ∈ Z.
(2) A perfectly generated triangulated category T has small products. In
fact, Brown representability implies that for any family of objects Xi in T the

functor i T (−, Xi ) is represented by an object in T .
5.2. Localization functors via Brown representability

The existence of localization functors is basically equivalent to the existence of
certain right adjoints; see Proposition 4.9.1. We combine this observation with
Brown’s representability theorem and obtain the following.
194 Henning Krause
Proposition 5.2.1. Let T be a triangulated category which admits small

coproducts and fix a localizing subcategory S.
(1) Suppose S is perfectly generated. Then there exists an exact localization

functor L : T → T with Ker L = S.
(2) Suppose T is perfectly generated. Then there exists an exact localization
functor L : T → T with Ker L = S if and only if the morphisms between
any two objects in T /S form a small set.
Proof. The existence of a localization functor L with Ker L = S is equivalent

to the existence of a right adjoint for the inclusion S → T , and it is equivalent to
the existence of a right adjoint for the quotient functor T → T /S. Both functors
preserve small coproducts since S is closed under taking small coproducts;
see Proposition 3.5.1. Now apply Theorem 5.1.1 for the existence of right
adjoints.
5.3. Compactly generated triangulated categories

Let T be a triangulated category with small coproducts. An object X in T is
%
called compact (or small) if every morphism X → i∈I Yi in T factors through
%
i∈J Yi for some finite subset J ⊆ I . Note that X is compact if and only if
the representable functor T (X, −) : T → Ab preserves small coproducts. The
compact objects in T form a thick subcategory which we denote by T c .
The triangulated category T is called compactly generated if it is perfectly
generated by a small set of compact objects. Observe that condition (PG2) is
automatically satisfied if every object in S is compact.
A compactly generated triangulated category T is perfectly cogenerated. To
see this, let S be a set of compact generators. Then the objects representing
HomZ (T (C, −), Q/Z), where C runs through all objects in S, form a set of
perfect cogenerators for T .
The following proposition is a reformulation of Brown representability for
compactly generated triangulated categories.
Proposition 5.3.1. Let F : T → U be an exact functor between triangulated

categories. Suppose that T has small coproducts and that T is compactly
generated.
(1) The functor F admits a right adjoint if and only if F preserves small
coproducts.
(2) The functor F admits a left adjoint if and only if F preserves small products.
5.4. Right adjoint functors preserving coproducts

The following lemma provides a characterization of the fact that a right adjoint
functor preserves small coproducts. This will be useful in the context of com-
pactly generated categories.
Lemma 5.4.1. Let F : T → U be an exact functor between triangulated cat-
egories which has a right adjoint G.
(1) If G preserves small coproducts, then F preserves compactness.
(2) If F preserves compactness and T is generated by compact objects, then
G preserves small coproducts.
Proof. Let X be an object in T and (Yi )i∈I a family of objects in U.
(1) We have
& & &
U(F X, Yi ) ∼
= T (X, G( Yi )) ∼ = T (X, GYi ). (5.4.1)
i i i
%
If X is compact, then the isomorphism shows that a morphism F X → i Yi
factors through a finite coproduct. It follows that F X is compact.
%
(2) Let X be compact. Then the canonical morphism φ :
% i GYi →
G( i Yi ) induces an isomorphism
& & & &
T (X, GYi ) ∼
= T (X, GYi ) ∼= U(F X, Yi ) ∼
= U(F X, Yi )
i i i i
&
∼
= T (X, G( Yi )),
i
where the last isomorphism uses that F X is compact. It is easily checked that
the objects X in T such that T (X , φ) is an isomorphism form a localizing
subcategory of T . Thus φ is an isomorphism because the compact objects
generate T .
5.5. Localization functors preserving coproducts

The following result provides a characterization of the fact that an exact local-
ization functor L preserves small coproducts; in that case one calls L smashing.
The example given below explains this terminology.
Proposition 5.5.1. Let T be a category with small coproducts and L : T → T
an exact localization functor. Then the following are equivalent.
(1) The functor L : T → T preserves small coproducts.
(2) The colocalization functor Γ : T → T with Ker Γ = Im L preserves small
coproducts.
196 Henning Krause
(3) The right adjoint of the inclusion functor Ker L → T preserves small
coproducts.
(4) The right adjoint of the quotient functor T → T / Ker L preserves small
coproducts.
(5) The subcategory Im L of all L-local objects is closed under taking small
coproducts.
If T is perfectly generated, in addition the following is equivalent.
(6) There exists a recollement of the following form.

o o
Im L o / T / Ker L (5.5.1)
inc
o
Proof. (1) ⇔ (4) ⇔ (5) follows from Proposition 2.7.1.
(1) ⇔ (2) ⇔ (3) is easily deduced from the functorial triangle (4.11.1)
relating L and Γ .
(5) ⇔ (6): Assume that T is perfectly generated. Then we can apply Brown’s
representability theorem and consider the sequence
I Q
Im L −→ T −→ Ker L
where I denotes the inclusion and Q a right adjoint of the inclusion Ker L → T .
∼
Note that Q induces an equivalence T / Im L − → Ker L; see Propositions 4.11.1
and 4.12.1. The functors I and Q have left adjoints. Thus the pair (I, Q) gives
rise to a recollement if and only if I and Q both admit right adjoints. It follows
from Proposition 4.9.1 that this happens if and only if Q admits a right adjoint.
Now Brown’s representability theorem implies that this is equivalent to the
fact that Q preserves small coproducts. And Proposition 3.5.1 shows that Q
preserves small coproducts if and only if Im L is closed under taking small
coproducts. This finishes the proof.
Remark 5.5.2. (1) The implication (6) ⇒ (5) holds without any extra assump-
tion on T .
(2) Suppose an exact localization functor L : T → T preserves small
coproducts. If T is compactly generated, then Im L is compactly generated. This
follows from Lemma 5.4.1, because the left adjoint of the inclusion Im L → T
sends the compact generators of T to compact generators for Im L. A similar
argument shows that Im L is perfectly generated provided that T is perfectly
generated.
Example 5.5.3. Let S be the stable homotopy category of spectra and ∧ its
smash product. Then an exact localization functor L : S → S preserves small
coproducts if and only if L is of the form L = − ∧ E for some spectrum E. We
sketch the argument. Let S denote the sphere spectrum. There exists a natural
morphism ηX : X ∧ LS → LX for each X in S. Suppose that L preserves
small coproducts. Then the subcategory of objects X in S such that ηX is
invertible contains S and is closed under forming small coproduts and exact
triangles. Thus L = − ∧ E for E = LS.
Let L : T → T be an exact localization functor which induces a recollement

of the form (5.5.1). Then the sequence Ker L → T → Im L of left adjoint
functors induces a sequence
(Ker L)c −→ T c −→ (Im L)c
of exact functors, by Lemma 5.4.1. This sequence is of some interest. The

functor (Ker L)c → T c is fully faithful and identifies (Ker L)c with a thick
subcategory of T c , whereas the functor T c → (Im L)c shares some formal
properties with a quotient functor. A typical example arises from finite local-
ization; see Theorem 5.6.1. However, there are examples where T is compactly
generated but 0 = (Ker L)c ⊆ Ker L = 0; see [25] for details.
5.6. Finite localization

A common type of localization for triangulated categories is finite localization.
Here, we explain the basic result and refer to our discussion of well generated
categories for a more general approach and further details.
Let T be a compactly generated triangulated category and suppose we have
given a subcategory S ⊆ T c . Let S denote the localizing subcategory generated
by S . Then S is compactly generated and therefore the inclusion functor S →
T admits a right adjoint by Brown’s representability theorem. In particular,
we have a localization functor L : T → T with Ker L = S and the morphisms
between any pair of objects in T /S form a small set; see Proposition 5.2.1.
We observe that the compact objects of S identify with the smallest thick
subcategory of T c containing S . This follows from Corollary 7.2.2. Thus we
obtain the following commutative diagram of exact functors.
Sc
inc /Tc can / T c /S c
inc inc J
I =inc
S /T Q=can
/ T /S
Theorem 5.6.1. Let T and S be as above. Then the quotient category T /S

is compactly generated. The induced exact functor J : T c /S c → T /S is fully
faithful and the category (T /S)c of compact objects equals the full subcategory
198 Henning Krause
consisting of all direct factors of objects in the image of J . Moreover, the

inclusion S ⊥ → T induces the following recollement.
o o
S⊥ o / T / S
inc
o
Proof. The inclusion I preserves compactness and therefore the right adjoint Iρ
preserves small coproducts by Lemma 5.4.1. Thus Qρ preserve small coprod-
ucts by Proposition 5.5.1, and therefore Q preserves compactness, again by
Lemma 5.4.1. It follows that J induces a functor T c /S c → (T /S)c . In partic-
ular, Q sends a set of compact generators of T to a set of compact generators
for T /S.
Next we apply Lemma 4.7.1 to show that J is fully faithful. For this, one
needs to check that every morphism from a compact object in T to an object
in S factors through some object in S c . This follows from Theorem 7.2.1.
The image of J is a full triangulated subcategory of T c which generates T /S.
Another application of Corollary 7.2.2 shows that every compact object of T /S
is a direct factor of some object in the image of J .
Let L : T → T denote the localization functor with Ker L = S. Then S ⊥
equals the full subcategory of L-local objects. This subcategory is closed under
small coproducts since S is generated by compact objects. Thus the existence
of the recollement follows from Proposition 5.5.1.
5.7. Cohomological localization via localization of graded modules

Let T be a triangulated category which admits small coproducts. Suppose that
T is generated by a small set of compact objects. We fix a graded1 ring and
a graded cohomological functor
H ∗ : T −→ A
into the category A of graded -modules. Thus H ∗ is a functor which sends

each exact triangle in T to an exact sequence in A, and we have an isomorphism
H∗ ◦S ∼ = T ◦ H ∗ where T denotes the shift functor for A. In addition, we
assume that H ∗ preserves small products and coproducts.
Theorem 5.7.1. Let L : A → A be an exact localization functor for the cate-

gory A of graded -modules. Then there exists an exact localization functor
1 All graded rings and modules are graded over Z. Morphisms between graded modules are
degree zero maps.
L̃ : T → T such that the following square commutes up to a natural isomor-

phism.
T
L̃ /T
H∗ H∗

A
L /A
More precisely, the adjunction morphisms Id A → L and Id T → L̃ induce for

each X in T the following isomorphisms.
∼ ∼
H ∗ L̃X −→ L(H ∗ L̃X) = LH ∗ (L̃X) ←− LH ∗ (X)
An object X in T is L̃-acyclic if and only if H ∗ X is L-acyclic. If an object X

in T is L̃-local, then H ∗ X is L-local. The converse holds, provided that H ∗
reflects isomorphisms.
Proof. We recall that T is perfectly cogenerated because it is compactly gen-

erated. Thus Brown’s representability theorem provides a compact object C in
T such that
&
H ∗X ∼= T (C, X)∗ := T (C, S i Y ) for all X ∈ T .
i∈Z
Now consider the essential image Im L of L which equals the full subcate-
gory formed by all L-local objects in A. Because L is exact, this subcategory
is coherent, that is, for any exact sequence X1 → X2 → X3 → X4 → X5 with
X1 , X2 , X4 , X5 ∈ A, we have X3 ∈ A. This is an immediate consequence of
the 5-lemma. In addition, Im L is closed under taking small products. The L-
local objects form an abelian Grothendieck category and therefore Im L admits
an injective cogenerator, say I ; see [16]. Using again Brown’s representability
theorem, there exists I˜ in T such that
A(H ∗ −, I ) ∼
= T (−, I˜) and therefore A(H ∗ −, I )∗ ∼
= T (−, I˜)∗ . (5.7.1)
Now consider the subcategory V of T which is formed by all objects X in T
such that H ∗ X is L-local. This is a triangulated subcategory which is closed
under taking small products. Observe that I˜ belongs to V. To prove this, take a
free presentation
F1 −→ F0 −→ H ∗ C −→ 0
over and apply A(−, I )∗ to it. Using the isomorphism (5.7.1), we see that
H ∗ I˜ belongs to Im L because Im L is coherent and closed under taking small
products.
200 Henning Krause
Now let U denote the smallest triangulated subcategory of T containing I˜

and closed under taking small products. Observe that U ⊆ V. We claim that U
is perfectly cogenerated by I˜. Thus, given a family of morphisms φi : Xi → Yi
in U such that T (Yi , I˜) → T (Xi , I˜) is surjective for all i, we need to show that

T ( i Yi , I˜) → T ( i Xi , I˜) is surjective. We argue as follows. If T (Yi , I˜) →
T (Xi , I˜) is surjective, then the isomorphism (5.7.1) implies that H ∗ φi is a
monomorphism since I is an injective cogenerator for Im L. Thus the product
∗
∗
i φi : iX i → i Yi induces a monomorphism H i φi = i H φi and

therefore T ( i φi , I˜) is surjective. We conclude from Brown’s representability
theorem that the inclusion functor G : U → T has a left adjoint F : T → U.
Thus L̃ = G ◦ F is a localization functor by Corollary 2.4.2.
Next we show that an object X ∈ T is L̃-acyclic if and only if H ∗ X is
L-acyclic. This follows from Proposition 4.10.1 and the isomorphism (5.7.1),
because we have
L̃X = 0 ⇐⇒ T (X, I˜) = 0 ⇐⇒ A(H ∗ X, I ) = 0 ⇐⇒ LH ∗ X = 0.
Now denote by η : Id A → L and η̃ : Id T → L̃ the adjunction morphisms
and consider the following commutative square.
ηH ∗ X
H ∗X / LH ∗ X (5.7.2)
H ∗ η̃X LH ∗ η̃X
∗
H ∗ L̃X
ηH L̃X
/ LH ∗ L̃X
We claim that LH ∗ η̃X and ηH ∗ L̃X are invertible for each X in T . The mor-
phism η̃X induces an exact triangle
η̃X
X → X −→ L̃X → SX
with L̃X = 0 = L̃SX . Applying the cohomological functor LH ∗ , we see that
LH ∗ η̃X is an isomorphism, since LH ∗ X = 0 = LH ∗ SX . Thus LH ∗ η̃ is
invertible. The morphism ηH ∗ L̃X is invertible because H ∗ L̃X is L-local. This
follows from the fact that L̃X belongs to U.
The commutative square (5.7.2) implies that H ∗ η̃X is invertible if and only
if ηH ∗ X is invertible. Thus if X is L̃-local, then H ∗ X is L-local. The converse
holds if H ∗ reflects isomorphisms.
Remark 5.7.2. (1) The localization functor L̃ is essentially uniquely deter-
mined by H ∗ and L, because Ker L̃ = Ker LH ∗ .
(2) Suppose that C is a generator of T . If L preserves small coproducts,
then it follows that L̃ preserves small coproducts. In fact, the assumption
implies that H ∗ L̃ preserves small coproducts, since LH ∗ ∼= H ∗ L̃. But H ∗

reflects isomorphisms because C is a generator of T . Thus L̃ preserves small
coproducts.
5.8. Example: Resolutions of chain complexes

Let A be an associative ring. Then the derived category D(A) of unbounded
chain complexes of modules over A is compactly generated. A compact
generator is the ring A, viewed as a complex concentrated in degree zero.
Let us be more precise, because we want to give an explicit construction
of D(A) which implies that the morphisms between any two objects in
D(A) form a small set. Moreover, we combine Brown representability with
Proposition 4.9.1 to provide descriptions of the adjoints Qλ and Qρ of the
quotient functor Q : K(Mod A) → D(A) which appear in the recollement
(4.14.1).
Denote by Loc A the localizing subcategory of K(Mod A) which is gen-
erated by A. Then Loc A is a compactly generated triangulated category and
(Loc A)⊥ = Kac (Mod A) since
K(Mod A)(A, S n X) ∼
= H n X.
Brown representability provides a right adjoint for the inclusion Loc A →

inc can
K(Mod A) and therefore the composite F : Loc A −→ K(Mod A) −→ D(A)
is an equivalence by Proposition 4.9.1. The right adjoint of the inclusion
Loc A → K(Mod A) annihilates the acyclic complexes and induces therefore
a functor D(A) → Loc A (which is a quasi-inverse for F ). The composite with
the inclusion Loc A → K(Mod A) is the left adjoint Qλ of Q and takes a
complex to its K-projective resolution.
Now fix an injective cogenerator I for the category of A-modules, for
instance I = HomZ (A, Q/Z). We denote by Coloc I the smallest thick subcat-
egory of K(Mod A) closed under small products and containing I . Then I is a
perfect cogenerator for Coloc I and ⊥ (Coloc I ) = Kac (Mod A) since
K(Mod A)(S n X, I ) ∼
= HomA (H n X, I ).
Brown representability provides a left adjoint for the inclusion Coloc I →

inc can
K(Mod A) and therefore the composite G : Coloc I −→ K(Mod A) −→ D(A)
is an equivalence by Proposition 4.9.1. The left adjoint of the inclusion
Coloc I → K(Mod A) annihilates the acyclic complexes and induces therefore
a functor D(A) → Coloc I (which is a quasi-inverse for G). The composition
202 Henning Krause
with the inclusion Coloc I → K(Mod A) is the right adjoint Qρ of Q and takes
a complex to its K-injective resolution.
5.9. Example: Homological epimorphisms

Let f : A → B be a ring homomorphism and f∗ : D(B) → D(A) the functor
given by restriction of scalars. Clearly, f∗ preserves small products and coprod-
ucts. Thus Brown representability implies the existence of left and right adjoints
for f∗ since D(A) is compactly generated. For instance, the left adjoint is the
derived tensor functor − ⊗LA B : D(A) → D(B) which preserves compactness.
The functor f∗ is fully faithful if and only if (f∗ −) ⊗LA B ∼= Id D(B) iff B ⊗A
B∼ = B and TorA i (B, B) = 0 for all i > 0. In that case f is called homological
epimorphism and the exact functor L : D(A) → D(A) sending X to f∗ (X ⊗LA B)
is a localization functor.
Take for instance a commutative ring A and let f : A → S −1 A = B be the
localization with respect to a multiplicatively closed subset S ⊆ A. Then the
induced exact localization functor L : D(A) → D(A) takes a chain complex X
to S −1 X. Note that L preserves small coproducts. In particular, L gives rise to
the following recollement.
o RHomA (B,−)
o
D(B) o / D(A) / U
f∗
o
−⊗LA B
The triangulated category U is equivalent to the kernel of − ⊗LA B, and one can
show that Ker(− ⊗LA B) is the localizing subcategory of D(A) generated by the
complexes of the form
x
··· → 0 → A −
→ A → 0 → ··· (x ∈ S).
5.10. Notes
The Brown representability theorem in homotopy theory is due to Brown
[9]. Generalizations of the Brown representability theorem for triangulated
categories can be found in work of Franke [13], Keller [21], and Neeman [31,
33]. The version used here is taken from [24]. The finite localization theorem
for compactly generated triangulated categories is due to Neeman [30]; it is
based on previous work of Bousfield, Ravenel, Thomason-Trobaugh, Yao, and
others. The cohomological localization functors commuting with localization
functors of graded modules have been used to set up local cohomology functors
in [7].
6. Well generated triangulated categories

6.1. Regular cardinals
A cardinal α is called regular if α is not the sum of fewer than α cardinals, all
smaller than α. For example, ℵ0 is regular because the sum of finitely many
finite cardinals is finite. Also, the successor κ + of every infinite cardinal κ
is regular. In particular, there are arbitrarily large regular cardinals. For more
details on regular cardinals, see for instance [26].
6.2. Localizing subcategories

Let T be a triangulated category and α a regular cardinal. A coproduct in T is
called α-coproduct if it has less than α factors. A full subcategory of T is called
α-localizing if it is a thick subcategory and closed under taking α-coproducts.
Given a subcategory S ⊆ T , we denote by Locα S the smallest α-localizing
subcategory of T which contains S. Note that Locα S is small provided that S
is small.
A full subcategory of T is called localizing if it is a thick subcategory
and closed under taking small coproducts. The smallest localizing subcategory

containing a subcategory S ⊆ T is Loc S = α Locα S where α runs through
all regular cardinals. We call Loc S the localizing subcategory generated by S.
6.3. Well generated triangulated categories

Let T be a triangulated category which admits small coproducts and fix a regular
%
cardinal α. An object X in T is called α-small if every morphism X → i∈I Yi
%
in T factors through i∈J Yi for some subset J ⊆ I with card J < α. The
triangulated category T is called α-well generated if it is perfectly generated
by a small set of α-small objects, and T is called well generated if it is β-well
generated for some regular cardinal β.
Suppose T is α-well generated by a small set S of α-small objects. Given
any regular cardinal β ≥ α, we denote by T β the β-localizing subcategory
Locβ S generated by S and call the objects of T β β-compact. Choosing a
representative for each isomorphism class, one can show that the β-compact
objects form a small set of β-small perfect generators for T . Moreover, T β
does not depend on the choice of S. For a proof we refer to [23, Lemma 5];

see also Proposition 6.8.1 and Remark 6.10.2. Note that T = β T β , where β

runs through all regular cardinals greater or equal than α, because β T β is a
triangulated subcategory containing S and closed under small coproducts.
204 Henning Krause
Remark 6.3.1. The α-small objects of T form an α-localizing subcategory.
Example 6.3.2. A triangulated category T is ℵ0 -well generated if and only if

T is compactly generated. In that case T ℵ0 = T c .
Example 6.3.3. Let A be the category of sheaves of abelian groups on a

non-compact, connected manifold of dimension at least 1. Then the derived
category D(A) of unbounded chain complexes is well generated, but the only
compact object in D(A) is the zero object; see [34]. For more examples of well
generated but not compactly generated triangulated categories, see [32].
6.4. Filtered categories

Let α be a regular cardinal. A category C is called α-filtered if the following
holds.
(FIL1) There exists an object in C.

(FIL2) For every family (Xi )i∈I of fewer than α objects there exists an object
X with morphisms Xi → X for all i.
(FIL3) For every family (φi : X → Y )i∈I of fewer than α morphisms there
exists a morphism ψ : Y → Z with ψφi = ψφj for all i and j .
One drops the cardinal α and calls C filtered in case it is ℵ0 -filtered.

Given a functor F : C → D, we use the term α-filtered colimit for the colimit
colim F X provided that C is a small α-filtered category.
−−−→
X∈C
Lemma 6.4.1. Let i : C → C be a fully faithful functor with C a small α-filtered

category. Suppose that i is cofinal in the sense that for any X ∈ C there is an
object Y ∈ C and a morphism X → iY . Then C is a small α-filtered category,
and for any functor F : C → D into a category which admits α-filtered colimits,
the natural morphism
colim F (iY ) −→ colim F X

−−−→
−−−→
Y ∈C X∈C
is an isomorphism.
Proof. See [19, Proposition 8.1.3].
A full subcategory C of a small α-filtered category C is called cofinal if for

any X ∈ C there is an object Y ∈ C and a morphism X → Y .
6.5. Comma categories

Let T be a triangulated category which admits small coproducts and fix a
full subcategory S. Given an object X in T , let S/X denote the category
whose objects are pairs (C, μ) with C ∈ S and μ ∈ T (C, X). The morphisms
(C, μ) → (C , μ ) are the morphisms γ : C → C in T making the following
C5
γ
/ C
55
μ 5
5
5 μ
X
Analogously, one defines for a morphism φ : X → X in T the category S/φ
whose objects are commuting squares in T of the form
C
γ
/ C
μ μ

X
φ
/ X
with C, C ∈ S.
Lemma 6.5.1. Let α be a regular cardinal and S an α-localizing subcategory

of T . Then the categories S/X and S/φ are α-filtered for each object X and
each morphism φ in T .
Proof. Straightforward.
6.6. The comma category of an exact triangle

Let T be a triangulated category. We consider the category of pairs (φ1 , φ2 ) of
φ1 φ2
composable morphisms X1 − → X2 − → X3 in T . A morphism μ : (φ1 , φ2 ) →
(φ1 , φ2 ) is a triple μ = (μ1 , μ2 , μ3 ) of morphisms in T making the following
X1
φ1
/ X2 φ2
/ X3
μ1 μ2 μ3
φ1 φ2
X1 / X / X
2 3
A pair (φ1 , φ2 ) of composable morphisms is called exact if it fits into an exact

φ1 φ2 φ3
triangle X1 −
→ X2 −
→ X3 −
→ SX1 .
206 Henning Krause
Lemma 6.6.1. Let μ : (γ1 , γ2 ) → (φ1 , φ2 ) be a morphism between pairs of

composable morphisms and suppose that (φ1 , φ2 ) is exact. Then μ factors
through an exact pair of composable morphisms which belong to the smallest
full triangulated subcategory containing γ1 and γ2 .
Proof. We proceed in two steps. The first step provides a factorization of μ

through a pair (γ1 , γ2 ) of composable morphisms such that γ2 γ1 = 0. To achieve
γ1 γ̄2
this, complete γ1 to an exact triangle C1 −
→ C2 −
→ C̄3 → SC1 . Note( that φ2 μ2
γ2 )
*γ + γ̄2
factors through γ̄2 . Now complete γ̄22 to an exact triangle C2 −−→ C3
[ δ δ̄ ]
C̄3 −−→ C3 → SC2 and observe that μ3 factors through δ via a morphism
μ3 : C3 → X3 . Thus we obtain the following factorization of μ with (δγ2 )γ1 =
−δ̄ γ̄2 γ1 = 0.
C1
γ1
/ C2 γ2
/ C3
id id δ

C1
γ1
/ C2 δγ2
/ C
3
μ1 μ2 μ3

X1
φ1
/ X2 φ2
/ X3
For the second step we may assume that γ2 γ1 = 0. We complete γ2 to an

γ̄1 γ2
exact triangle C̄1 −
→ C2 − → C3 → S C̄1 . Clearly, γ1 factors through γ̄1 via a
morphism ρ : C1 → C̄1 and μ2 γ̄1 factors through φ1 via a morphism σ : C̄1 →
X1 . Thus we obtain the following factorization of μ
C1
γ1
/ C2 γ2
/ C3
( ) ( )
ρ id
id id ( ) 0
γ2

C̄1 C1
[ γ̄1 0 ]
/ C2 0
/ C3 SC1
[ σ μ1 −σρ ] μ2 [ μ3 0 ]

X1
φ1
/ X2 φ2
/ X3
where the middle row fits into an exact triangle.
The following statement is a reformulation of the previous one in terms of

cofinal subcategories.
Proposition 6.6.2. Let T be a triangulated category and S a full triangulated

φ1 φ2 φ3
subcategory. Suppose that X1 − → X2 − → X3 − → SX1 is an exact triangle in
T and denote by S/(φ1 , φ2 ) the category whose objects are the commutative
diagrams in T of the following form.
C1
γ1
/ C2 γ2
/ C3

X1
φ1
/ X2 φ2
/ X3
such each Ci belongs to S. Then the full subcategory formed by the diagrams
γ1 γ2 γ3
such that there exists an exact triangle C1 −
→ C2 −→ C3 − → SC1 is a cofinal
subcategory of S/(φ1 , φ2 ).
6.7. A Kan extension

Let T be a triangulated category with small coproducts and S a small full sub-
category. Suppose that the objects of S are α-small and that S is closed under α-
coproducts. We denote by Addα (S op , Ab) the category of α-product preserving
functors S op → Ab. This is a locally presentable abelian category in the sense
of [17] and we refer to the Appendix B for basic facts on locally presentable
categories. Depending on the choice of S, we can think of Addα (S op , Ab) as a
locally presentable approximation of the triangulated category T . In order to
make this precise, we need to introduce various functors.
Let hT : T → A(T ) denote the abelianization of T ; see Appendix A. Some-
times we write T instead of A(T ). The inclusion functor f : S → T induces a
functor
f∗ : A(T ) −→ Addα (S op , Ab), X
→ A(T )((hT ◦ f )−, X),
and we observe that the composite
hT f∗
T −→ A(T ) −→ Addα (S op , Ab)
is the restricted Yoneda functor sending each X ∈ T to T (−, X)|S .
The next proposition discusses a left adjoint for f∗ . To this end, we denote
for any category C by hC the Yoneda functor sending X in C to C(−, X).
Proposition 6.7.1. The functor f∗ admits a left adjoint f ∗ which makes the
following diagram commutative.
S
hS
/ Addα (S op , Ab)
f =inc f∗

T
hT
/ A(T )
208 Henning Krause
Moreover, the functor f ∗ has the following properties.
(1) f ∗ is fully faithful and identifies Addα (S op , Ab) with the full subcategory
formed by all colimits of objects in {T (−, X) | X ∈ S}.
(2) f∗ preserves small coproducts if and only if (PG2) holds for S.
(3) Suppose that S is a triangulated subcategory of T . Then for X in T , the
adjunction morphism f ∗ f∗ (hT X) → hT X identifies with the canonical
morphism
colim hT C −→ hT X.
−−−→
(C,μ)∈S/X
Proof. The functor f ∗ is constructed as a left Kan extension. To explain this,

it is convenient to identify Addα (S op , Ab) with the category Lexα (Sop , Ab) of
left exact functors Sop → Ab which preserve α-products. To be more precise,
the Yoneda functor h : S → S induces an equivalence
∼
Lexα (Sop , Ab) −→ Addα (S op , Ab), F
→ F ◦ h,
because every additive functor S op → Ab extends uniquely to a left exact

functor Sop → Ab; see Lemma A.1.
Using this identification, the existence of a fully faithful left adjoint
Lexα (Sop , Ab) → A(T ) for f∗ and its basic properties follow from Lemma B.6,
because the inclusion f : S → T induces a fully faithful and right exact func-
tor f: S → T = A(T ). This functor preserves α-coproducts and identifies S
with a full subcategory of α-presentable objects, since the objects from S are
α-small in T .
(2) Let = (f∗ ) denote the set of morphisms of A(T ) which f∗ makes
invertible. It follows from Proposition 2.3.1 that f∗ induces an equivalence
∼
A(T )[ −1 ] −→ Lexα (Sop , Ab),
and therefore f∗ preserves small coproducts if and only if is closed under

taking small coproducts, by Proposition 3.5.1. It is not hard to check that f∗ is
exact, and therefore a morphism in A(T ) belongs to if and only if its kernel
and cokernel are annihilated by f∗ . Now observe that an object F in A(T ) with
presentation T (−, X) → T (−, Y ) → F → 0 is annihilated by f∗ if and only
if T (C, X) → T (C, Y ) is surjective for all C ∈ S. It follows hat f∗ preserves
small coproducts if and only if (PG2) holds for S.
(3) Let F = f∗ (hT X) = T (−, X)|S . Then Lemma B.7 implies that F =
colim hS C, since S/F = S/X. Thus f ∗ F = colim hT C.
−−−→ −−−→
(C,μ)∈S/X (C,μ)∈S/X
Corollary 6.7.2. Let T be a triangulated category with small coproducts.

Suppose T is α-well generated and denote by T α the full subcategory formed
by all α-compact objects. Then the functor T → A(T ) taking an object X to
colim T (−, C)
−−−→
(C,μ)∈T α /X
preserves small coproducts.
6.8. A criterion for well generatedness

Let T be a triangulated category which admits small coproducts. The follow-
ing result provides a useful criterion for T to be well generated in terms of
cohomological functors into locally presentable abelian categories.
Proposition 6.8.1. Let T be a triangulated category with small coproducts and

α a regular cardinal. Let S0 be a small set of objects and denote by S the full
subcategory formed by all α-coproducts of objects in S0 . Then the following
are equivalent.
(1) The objects of S0 are α-small and (PG2) holds for S0 .

(2) The objects of S are α-small and (PG2) holds for S.
(3) The functor H : T → Addα (S op , Ab) taking X to T (−, X)|S preserves
small coproducts.
Proof. It is clear that (1) and (2) are equivalent, and it follows from Proposi-
tion 6.7.1 that (2) implies (3). To prove that (3) implies (2), assume that H pre-
%
serves small coproducts. Let φ : X → i∈I Yi be a morphism in T with X ∈ S.
% %
Write i∈I Yi = colim YJ as α-filtered colimit of coproducts YJ = i∈J Yi
−−−→
J ⊆I
with card J < α. Then we have
colim T (X, YJ ) ∼
= colim HomS (S(−, X), H YJ )
−−−→ −−−→
J ⊆I J ⊆I
∼
= HomS (S(−, X), colim H YJ )
−−−→
J ⊆I
&
∼
= HomS (S(−, X), H Yi )
i∈I
&
∼
= HomS (S(−, X), H ( Yi ))
i∈I
&
∼
= T (X, Yi ).
i∈I
210 Henning Krause
Thus φ factors through some YJ , and it follows that X is α-small. Now Propo-
sition 6.7.1 implies that (PG2) holds for S.
6.9. Cohomological functors via filtered colimits

The following theorem shows that cohomological functors on well generated
triangulated categories can be computed via filtered colimits. This generalizes
a fact which is well known for compactly generated triangulated categories. We
say that an abelian category has exact α-filtered colimits provided that every
α-filtered colimit of exact sequences is exact.
Theorem 6.9.1. Let T be a triangulated category with small coproducts.

Suppose T is α-well generated and denote by T α the full subcategory formed
by all α-compact objects. Let A be an abelian category which has small
coproducts and exact α-filtered colimits. If H : T → A is a cohomological
functor which preserves small coproducts, then we have for X in T a natural
isomorphism
∼
colim H C −→ H X. (6.9.1)
−−−→α
(C,μ)∈T /X
Proof. The left hand term of (6.9.1) defines a functor H̃ : T → A and we need
to show that the canonical morphism H̃ → H is invertible.
First observe that H̃ is cohomological. This is a consequence of Proposi-
tion 6.6.2 and Lemma 6.4.1, because for any exact triangle X1 → X2 → X3 →
SX1 in T , the sequence H̃ X1 → H̃ X2 → H̃ X3 can be written as α-filtered
colimit of exact sequences in A.
Next we claim that H̃ preserves small coproducts. To this end consider the
exact functor H̄ : A(T ) → A which extends H ; see Lemma A.2. Note that H̄
preserve small coproducts because H has this property. We have for X in T

H̃ X = colim H̄ T (−, C) ∼ = H̄ colim T (−, C) .
−−−→α −−−→α
(C,μ)∈T /X (C,μ)∈T /X
Now the assertion follows from Corollary 6.7.2.

To complete the proof, consider the full subcategory T consisting of those
objects X in T such that the morphism H̃ X → H X is an isomorphism. Clearly,
T is a triangulated subcategory since both functors are cohomological, it is
closed under taking small coproducts since they are preserved by both functors,
and it contains T α . Thus T = T .
Remark 6.9.2. For an alternative proof of the fact that H̃ is cohomological,

one uses Lemma B.5.
6.10. A universal property

Let T be a triangulated category which admits small coproducts and is α-well
generated. We denote by Aα (T ) the full subcategory of A(T ) which is formed
by all colimits of objects T (−, X) with X in T α . Observe that Aα (T ) is a locally
presentable abelian category with exact α-filtered colimits. This follows from
Proposition 6.7.1 and the discussion in Appendix B, because Aα (T ) can be
identified with a category of left exact functors.
We have two functors
Hα : T −→ Aα (T ), X
→ colim T (−, C),
−−−→
(C,μ)∈T α /X
hα : T −→ Addα ((T α )op , Ab), X
→ T (−, X)|T α ,
which are related by an equivalence as follows.
α op
2 Addα ((T ) , Ab)
eeeeeeeeeee
hα
T YeYYYYYYY ∼ f∗
YYYYYY
Hα YY,
Aα (T )
The functor f ∗ is induced by the inclusion f : T α → T and discussed in

Proposition 6.7.1. In particular, there it is shown that f ∗ (hα X) = f ∗ f∗ (hT X) =
Hα X for all X in T .
Proposition 6.10.1. The functor Hα : T → Aα (T ) has the following universal

property.
(1) The functor Hα is a cohomological functor to an abelian category with

small coproducts and exact α-filtered colimits and Hα preserves small
coproducts.
(2) Given a cohomological functor H : T → A to an abelian category with
small coproducts and exact α-filtered colimits such that H preserves small
coproducts, there exists an essentially unique exact functor H̄ : Aα (T ) →
A which preserves small coproducts and satisfies H = H̄ ◦ Hα .
Proof. (1) It is clear that hα is cohomological and it follows from Proposi-

tion 6.7.1 that hα preserves small coproducts.
(2) Given H : T → A, we denote by H̃ : A(T ) → A the exact functor
which extends H , and we define H̄ : Aα (T ) → A by sending each X to H̃ X.
212 Henning Krause
The following commutative diagram illustrates this construction.
hA(T α )
Tα
hT α
/ A(T α ) / Aα (T )
f =inc A(f ) f ∗ =inc

T
hT
/ A(T ) A(T )
H̃

T
H /A
Let us check the properties of H̄ . The functor H̄ preserves small coproducts

since H̃ has this property. The functor H̄ is exact when restricted to A(T α ).
Thus it follows from Lemma B.5 that H̄ is exact. The equality H = H̄ ◦ Hα is
a consequence of Theorem 6.9.1 since both functors coincide on T α . Suppose
now there is a second functor Aα (T ) → A having the properties of H̄ . Then
both functors agree on P = {T (−, X) | X ∈ T α } and therefore on all of Aα (T )
since each object in Aα (T ) is a colimit of objects in P and both functors preserve
colimits.
Remark 6.10.2. The universal property can be used to show that the category
T α of α-compact objects does not depend on the choice of a perfectly generating
set for T . More precisely, if T is α-well generated, then two α-localizing
subcategories coincide if each contains a small set of α-small perfect generators.
This follows from the fact that the functor Hα identifies the α-compact objects
with the α-presentable projective objects of Aα (T ).
6.11. Notes
Well generated triangulated were introduced and studied by Neeman in his
book [33] as a natural generalization of compactly generated triangulated cat-
egories. For an alternative approach which simplifies the definition, see [23].
More recently, well generated categories with specific models have been stud-
ied; see [37, 47] for work involving algebraic models via differential graded
categories, and [20] for topological models. In [43], Rosický used combina-
torial models and showed that there exist universal cohomological functors
into locally presentable categories which are full. Interesting consequences of
this fact are discussed in [35]. The description of the universal cohomological
functors in terms of filtered colimits seems to be new.
7. Localization for well generated categories

7.1. Cohomological localization
The following theorem shows that cohomological functors on well generated
triangulated categories induce localization functors. This generalizes a fact
which is well known for compactly generated triangulated categories.
Theorem 7.1.1. Let T be a triangulated category with small coproducts which

is well generated. Let H : T → A be a cohomological functor into an abelian
category which has small coproducts and exact α-filtered colimits for some
regular cardinal α. Suppose also that H preserves small coproducts. Then
there exists an exact localization functor L : T → T such that for each object
X we have LX = 0 if and only if H (S n X) = 0 for all n ∈ Z.
Proof. We may assume that T is α-well generated. Let = (H ) denote

the set of morphisms σ in T such H σ is invertible. Next we assume that
S = . Otherwise, we replace A by a countable product AZ of copies
of A and H by (H S n )n∈Z . Then admits a calculus of right fractions by
Lemma 4.4.2, and we apply the criterion of Lemma 3.3.1 to show that the
morphisms between any two objects in T [ −1 ] form a small set. The existence
of a localization functor L : T → T with Ker L = Ker H then follows from
Proposition 5.2.1.
Thus we need to specify for each object Y of T a small set of objects S(Y, )
such that for every morphism X → Y in , there exist a morphism X → X in
with X in S(Y, ). Suppose that Y belongs to T κ . We define by induction
κ−1 = κ + α and
κn = sup{card T α /U | U ∈ T κn−1 }+ + κn−1 for n ≥ 0.

Then we put S(Y, ) = T κ̄ with κ̄ = ( n≥0 κn )+ .
Now fix σ : X → Y in . The morphism X → X in with X in S(Y, )
%
is constructed as follows. The canonical morphism π : (C,μ)∈T α /X C → X
induces an epimorphism H π by Theorem 6.9.1. We can choose C ⊆ T α /X with
%
card C ≤ card T α /Y such that π0 : X0 = (C,μ)∈C C → X induces an epimor-
phism H π0 since H σ is invertible. More precisely, we call two objects (C, μ)
and (C , μ ) of T α /X equivalent if σ μ = σ μ , and we choose as objects of C
precisely one representative for each equivalence class.
Suppose we have already constructed πi : Xi → X with Xi in T κi for
some i ≥ 0. Then we form the following commutative diagram with exact
rows.
214 Henning Krause
Ui
ιi
/ Xi πi
/X / SUi
σi σ Sσi

Vi / Xi /Y / SVi
Note that H σi is invertible. Thus we can choose Ci ⊆ T α /Ui with

card Ci ≤ card T α /Vi ≤ card T α /Xi + card T α /Y < κi+1
%
such that ξi : (C,μ)∈Ci C → Ui induces an epimorphism H ξi . Now complete
ιi ◦ ξi to an exact triangle and define πi+1 : Xi+1 → X by the commutativity of
the following diagram.
% ιi ◦ ξi %
C / Xi φi
/ Xi+1 /S C
(C,μ)∈Ci 55
(C,μ)∈Ci
55
πi 55
5 πi+1
X
Observe that Xi+1 belongs to T κi+1 and that Ker H πi = Ker H φi . The φi induce
an exact triangle
& (id −φi ) & ψ &
Xi −−−−→ Xi −→ X −→ S Xi (7.1.1)
i∈N i∈N i∈N

%
such that X belongs to S(Y, ) and the morphism (πi ) : i∈N Xi → X factors
through ψ via a morphism τ : X → X. We claim that H τ is invertible. In fact,
the lemma below implies that the πi induce the following exact sequence
& (id −H φi ) & (H πi )
0 −→ H Xi −−−−−→ H Xi −−−→ H X −→ 0.
i∈N i∈N
On the other hand, the exact triangle (7.1.1) induces the exact sequence
& H (id −φi ) & Hψ &
H Xi −−−−−→ H Xi −→ H X −→ H S Xi ,
i∈N i∈N i∈N
and a comparison shows that H τ is invertible. Here, we use again that H

preserves small coproducts, and this completes the proof.
Lemma 7.1.2. Let A be an abelian category which admits countable coprod-
ucts. Then a sequence of epimorphisms (πi )i∈N
Xi 6
φi
/ Xi+1
66
6
πi 6
6 i+1
π
Y
satisfying πi = πi+1 ◦ φi and Ker πi = Ker φi for all i induces an exact

sequence
& (id −φi ) & (πi )
0 −→ Xi −−−−→ Xi −−→ Y −→ 0.
i∈N i∈N
Proof. The assumption Ui := Ker πi = Ker φi implies that there exists a mor-
phism πi : Y → Xi with πi πi = id Y and φi πi = πi+1

for all i ≥ 1. Thus we
have a sequence of commuting squares
[ inc πi ] / Xi
Ui Y
( )
0 0 φi
0 id

[ inc πi+1 ]
Ui+1 Y / Xi+1
where the horizontal maps are isomorphisms. Taking colimits on both sides,
the assertion follows.
7.2. Localization with respect to a small set of objects

Let T be a well generated triangulated category and S a localizing subcategory
which is generated by a small set of objects. The following result says that

S and T /S are both well generated and that the filtration T = α T α via
α-compact objects induces canonical filtrations

S= (S ∩ T α ) and T /S = T α /(S ∩ T α ).
α α
Theorem 7.2.1. Let T be a well generated triangulated category and S a

localizing subcategory which is generated by a small set of objects. Fix a
regular cardinal α such that T is α-well generated and S is generated by
α-compact objects.
(1) An object X in T belongs to S if and only if every morphism C → X from

an object C in T α factors through some object in S ∩ T α .
(2) The localizing subcategory S and the quotient category T /S are α-well
generated.
(3) We have S α = S ∩ T α and a commutative diagram of exact functors
Sα
inc /Tα can / T α /S α
inc inc J

S
inc /T can / T /S
216 Henning Krause
such that J is fully faithful. Moreover, J induces a functor T α /S α →

(T /S)α such that every object of (T /S)α is a direct factor of an object in
the image of J . This functor is an equivalence if α > ℵ0 .
Proof. Let C = S ∩ T α . Then the inclusion i : C → T α induces a fully faith-

ful and exact functor i ∗ : Addα (C op , Ab) → Addα ((T α )op , Ab) which is left
adjoint to the functor i∗ taking F to F ◦ i; see Lemma B.8. Note that the image
Im i ∗ of i ∗ is closed under small coproducts. We consider the restricted Yoneda
functor hα : T → Addα ((T α )op , Ab) taking X to T (−, X)|T α and observe that
h−1 ∗
α (Im i ) is a localizing subcategory of T containing C. Thus we obtain a
functor H making the following diagram commutative.
S
inc /T
H hα
∗
Addα (C op , Ab)
i / Addα ((T α )op , Ab)
Let us compare H with the restricted Yoneda functor
H : S −→ Addα (C op , Ab), X
→ S(−, X)|C .
In fact, we have an isomorphism

∼
→ i∗ ◦ i ∗ ◦ H = i∗ ◦ hα |S = H
H −
and H preserves small coproducts since hα does. It follows from Proposi-

tion 6.8.1 that C provides a small set of α-small perfect generators for S. Thus
S is α-well generated and S α = Locα C = S ∩ T α .
Next we apply Proposition 5.2.1 and obtain a localization functor L : T →
T with Ker L = S. We use L to show that S = h−1 ∗
α (Im i ). We know already
that S ⊆ hα (Im i ). Now let X be an object in hα (Im i ∗ ) and consider the
−1 ∗ −1
exact triangle Γ X → X → LX → S(Γ X). Then T (C, LX) = 0 for all C ∈ C

and therefore i∗ hα LX = 0. On the other hand, hα LX = i ∗ F for some functor
F and therefore 0 = i∗ hα LX = i∗ i ∗ F ∼ = F . Thus LX = 0 and therefore X
−1 ∗
belongs to S. This shows S = hα (Im i ).
Now we prove (1) and use the description of the essential image of i ∗ from
Lemma B.8. We have for an object X in T that X belongs to S iff hα X belongs
to Im i ∗ iff every morphism T α (−, C) → hα X with C ∈ T α factors through
T α (−, C ) for some C ∈ C iff every morphism C → X with C ∈ T α factors
through some C ∈ C.
An immediate consequence of (1) is the fact that J is fully faithful. This
follows from Lemma 4.7.1.
Now consider the quotient functor q : T α → T α /S α . This induces an exact

functor q ∗ : Addα ((T α )op , Ab) → Addα ((T α /S α )op , Ab) which is left adjoint
to the fully faithful functor q∗ taking F to F ◦ q; see Lemma B.8. Clearly,
q ∗ ◦ hα annihilates S and induces therefore a functor K making the following
T
Q=can
/ T /S
hα K
q∗
Addα ((T α )op , Ab) / Addα ((T α /S α )op , Ab)
Note that Q admits a right adjoint which we denote by Qρ . We identify T α /S α

via J with a full triangulated subcategory of T /S and consider the restricted
Yoneda functor
K : T /S −→ Addα ((T α /S α )op , Ab), X

→ T /S(−, X)|T α /S α .
Adjointness gives the following isomorphism
T /S(J qC, X) = T /S(QC, X) ∼

= T (C, Qρ X)
for all C ∈ T α and X ∈ T /S. Thus we have an isomorphism
∼ ∼
− K ◦ Q ◦ Qρ = q ∗ ◦ hα ◦ Qρ ∼
K← = q ∗ ◦ q∗ ◦ K −
→ K
and K preserves small coproducts since hα does. It follows from Proposi-

tion 6.8.1 that T α /S α provides a small set of α-small perfect generators for
T /S. Thus T /S is α-well generated and (T /S)α = Locα (T α /S α ).
Corollary 7.2.2. Let T be an α-well generated triangulated category and S a

localizing subcategory generated by a small set S0 of α-compact objects. Then
S is α-well generated and S α equals the α-localizing subcategory generated
by S0 .
Proof. In the preceding proof of Theorem 7.2.1, we can choose for C instead
of S ∩ T α the α-localizing subcategory of T which is generated by S0 . Then
the proof shows that C provides a small set of α-small perfect generators for S.
Thus we have S α = C by definition.
The localization with respect to a localizing subcategory generated by a

small set of objects can be interpreted in various ways. The following remark
provides some indication.
Remark 7.2.3. (1) Let T be a well generated triangulated category and φ

a morphism in T . Then there exists a universal exact localization functor
218 Henning Krause
φ
L : T → T inverting φ. To see this, complete φ to an exact triangle X − →
Y → Z → SX and let L be the localization functor such that Ker L equals
the localizing subcategory generated by Z. Conversely, any exact localization
functor L : T → T is the universal exact localization functor inverting some
morphism φ provided that Ker L is generated by a small set S0 of objects. To
%
see this, take φ : 0 → X∈S0 X.
(2) Let T be a triangulated category and L : T → T an exact localization
functor such that S = Ker L is generated by a single object W . Then the
first morphism Γ X → X from the functorial triangle Γ X → X → LX →
S(Γ X) is called cellularization and the second morphism X → LX is called
nullification with respect to W . The objects in S are built from W .
7.3. Functors between well generated categories

We consider functors between well generated triangulated categories which
are exact and preserve small coproducts. The following result shows that such
functors are controlled by their restriction to the subcategory of α-compact
objects for some regular cardinal α.
Proposition 7.3.1. Let F : T → U be an exact functor between α-well gener-

ated triangulated categories. Suppose that F preserves small coproducts and
let G be a right adjoint.
(1) There exists a regular cardinal β0 ≥ α such that F preserves β0 -

compactness. In that case F preserves β-compactness for all regular
β ≥ β0 .
(2) Given a regular cardinal β ≥ β0 , the restriction f : T β → U β of F induces
the following diagram of functors which commute up to natural isomor-
phisms.
f =F β
Tβ / Uβ
inc inc

T
F /U G
/T
hβ (T ) hβ (U) hβ (T )
∗
Addβ ((T β )op , Ab)
f
/ Addβ ((U β )op , Ab) f∗
/ Addβ ((T β )op , Ab)
Proof. (1) Choose β0 ≥ α such that F (T α ) ⊆ U β0 . Then we get for β ≥ β0
F (T β ) = F (Locβ T α ) ⊆ Locβ F (T α ) ⊆ Locβ U β0 = U β .

(2) We apply Theorem 6.9.1 to show that hβ (U) ◦ F ∼

= f ∗ ◦ hβ (T ). In fact,
it follows from Proposition 6.10.1 and Lemma B.8 that both composites are
cohomological functors, preserve small coproducts, and agree on T β .
The isomorphism hβ (T ) ◦ G ∼
= f∗ ◦ hβ (U) follows from the adjointness of
F and G, since T (C, GX) ∼ = U(f C, X) for every C ∈ T β and X ∈ U.
7.4. The kernel of a functor between well generated categories

We show that the class of well generated triangulated categories is closed under
taking kernels of exact functors which preserve small coproducts.
Theorem 7.4.1. Let F : T → U be an exact functor between α-well generated
triangulated categories and suppose that F preserves small coproducts. Let
S = Ker F and choose a regular cardinal β ≥ α such that F preserves β-
compactness.
(1) An object X in T belongs to S if and only if every morphism C → X with
C ∈ T β factors through a morphism γ : C → C in T β satisfying F γ = 0.
(2) Suppose β > ℵ0 . Then S is β-well generated and S β = S ∩ T β .
Proof. Let f : T β → U β be the restriction of F and denote by I the set of
morphisms in T β which are annihilated by F .
(1) Let X be an object in T . Then it follows from Proposition 7.3.1 that
F X = 0 if and only f ∗ hβ X = 0. Now Lemma B.8 implies that f ∗ hβ X = 0 iff
each morphism C → X with C ∈ T β factors through some morphism C → C
in I.
(2) Let S denote the localizing subcategory of T which is generated by all
homotopy colimits of sequences
C0 −→ C1 −→ C2 −→ · · ·
of morphisms in I. We claim that S = Ker F . Clearly, we have S ⊆ Ker F .
Now fix an object X ∈ Ker F . We have seen in (1) that each morphism μ : C →
X with C ∈ T β factors through some morphism C → C in I. We obtain by
induction a sequence
γ0 γ1 γ2
C = C0 −→ C1 −→ C2 −→ · · ·
of morphisms in I such that μ factors through each finite composite γi . . . γ0 .
Thus μ factors through the homotopy colimit of this sequence and therefore
through an object of S ∩ T β . Here one uses that β > ℵ0 . We conclude from
Theorem 7.2.1 that X belongs to S . Moreover, we conclude from this theorem
that S is β-well generated.
220 Henning Krause
Remark 7.4.2. It is necessary to assume in part (2) of the preceding theorem

that β > ℵ0 . For example, there exists a ring A with Jacobson radical r such
that the functor F = − ⊗LA A/r : D(A) → D(A/r) satisfies S = Ker F = 0 but
S ∩ D(A)c = 0; see [22].
Observe that Theorem 7.4.1 provides a partial answer to the telescope con-
jecture for compactly generated categories. This conjecture claims that the
kernel of a localization functor L : T → T is generated by compact objects
provided that L preserves small coproducts. Part (1) implies that S = Ker L
is generated by morphisms between compact objects, and part (2) says that
S is generated by ℵ1 -compact objects. I am grateful to Amnon Neeman for
explaining to me how to deduce (2) from (1). The following corollary makes
the connection with the telescope conjecture more precise; just put α = ℵ0 .
Corollary 7.4.3. Let L : T → T be an exact localization functor which

preserves small coproducts. Suppose that T is α-well generated and let
β ≥ max(α, ℵ1 ). Then S = Ker L is β-well generated and S β = S ∩ T β .
Proof. Let L : T → T be a localization functor which preserves small coprod-

ucts. Write L = G ◦ F as the composite of a quotient functor F : T → U
and a fully faithful right adjoint G, where U = T /S and S = Ker L. Then G
preserves small coproducts by Proposition 5.5.1. The isomorphism (5.4.1)
from the proof of Lemma 5.4.1 shows that F preserves α-smallness and
sends a set of perfect generators of T to a set of perfect generators of U.
In particular, F preserves β-compactness for all regular β ≥ α. Now apply
Theorem 7.4.1.
7.5. The kernel of a cohomological functor on a well

generated category
The following result says that kernels of cohomological functors from well
generated triangulated categories into locally presentable abelian categories
are well generated. The argument is basically the same as that for kernels of
exact functors between well generated triangulated categories.
Theorem 7.5.1. Let H : T → A be a cohomological functor from a well

generated triangulated category into a locally presentable abelian category
and suppose that H preserves small coproducts. Let S denote the localizing
subcategory of T consisting of all objects X such that H (S n X) = 0 for all
n ∈ Z. Then S is a well generated triangulated category.
Proof. Replacing H by (H S n )n∈Z , we may assume that S = Ker H . Choose

a regular cardinal α such that T is α-well generated and A is locally
α-presentable. Then we have H (T α ) ⊆ Aβ for some regular cardinal β
and we assume β ≥ max(α, ℵ1 ). The description of H in Theorem 6.9.1
shows that H restricts to a functor h : T β → Aβ , and we denote by
h̄ : A(T β ) → Aβ the induced exact functor. Then we obtain the following
functor
∼ h̄∗ ∼
h∗ : Addβ ((T β )op , Ab) −
→ Lexβ (A(T β )op , Ab) −
→ Lexβ ((Aβ )op , Ab) −
→A
where the first equivalence follows from Lemma B.1 and the second equivalence
follows from Lemma B.6. The functor h̄∗ is a left Kan extension; it takes a
filtered colimit
F = colim A(T β )(−, C) to colim Aβ (−, h̄C).

−−−→ −−−→
(C,μ)∈A(T β )/F (C,μ)∈A(T β )/F
Note that h∗ is exact and preserves small coproducts. This follows from
Lemma B.5 and the fact that h̄∗ is left adjoint to the restriction functor h̄∗ .
The composite h∗ ◦ hβ : T → A coincides with H on T β and therefore
h ∗ ◦ hβ ∼
= H by Theorem 6.9.1. In particular, we have for each X in T that
H X = 0 if and only if h∗ (hβ X) = 0. Now we use the same argument as in
the proof of Theorem 7.4.1 and show that Ker H is generated by all homotopy
colimits of countable sequences of morphisms in T β which are annihilated by
H.
7.6. Localization of well generated categories versus

abelian localization
We demonstrate the interplay between triangulated and abelian localization. To
this end recall from Proposition 6.10.1 that we have for each well generated
category T a universal cohomological functor Hα : T → Aα (T ) into a locally
α-presentable abelian category. We show that each exact localization functor
for T can be extended to an exact localization functor for Aα (T ) for some
regular cardinal α.
Theorem 7.6.1. Let T be a well generated triangulated category and

L : T → T an exact localization functor. Suppose that Ker L is well generated.
Then there exists a regular cardinal α and an exact localization functor
222 Henning Krause
L : Aα (T ) → Aα (T ) such that the following square commutes up to a nat-

ural isomorphism.
T
L /T
Hα Hα
L

Aα (T ) / Aα (T )
More precisely, the adjunction morphisms Id T → L and Id Aα (T ) → L

induce for each X in T the following isomorphisms.
∼ ∼
Hα LX −→ L (Hα LX) = L Hα (LX) ←− L Hα (X)
An object X in T is L-acyclic if and only if Hα X is L -acyclic, and X is L-local
if and only if Hα X is L -local.
Proof. Choose a regular cardinal α > ℵ0 such that T is α-well generated
and S = Ker L is generated by α-compact objects. Let U = T /S and write
L = G ◦ F as the composite of the quotient functor F : T → U with its right
adjoint G : U → T .
Now identify Aα (T ) = Addα ((T α )op , Ab) and Aα (U ) = Addα ((U α )op , Ab).
The induced functor f : T α → U α equals, up to an equivalence, the quo-
tient functor T α → T α /S α , by Theorem 7.2.1. From f we obtain a pair of
adjoint functors f ∗ and f∗ by Lemma B.8. Both functors are exact and the
right adjoint f∗ is fully faithful. Thus we obtain an exact localization func-
tor L = f∗ ◦ f ∗ for Aα (T ) by Corollary 2.4.2. The commutativity Hα ◦ L ∼ =
L ◦ Hα and the assertions about acyclic and local objects then follow from
Proposition 7.3.1.
7.7. Example: The derived category of an abelian

Grothendieck category
Let A be an abelian Grothendieck category. Then the derived category D(A)
of unbounded chain complexes is a well generated triangulated category. Let
us sketch an argument.
The Popescu-Gabriel theorem says that for each generator G of A, the
functor T = A(G, −) : A → Mod A (where A = A(G, G) denotes the endo-
morphism ring of G) is fully faithful and admits an exact left adjoint, say Q;
see [45, Theorem X.4.1]. Consider the cohomological functor H : D(A) → A
%
taking a complex X to Q( n∈Z H n X). Then an application of Theorem 7.5.1
shows that S = Ker H is well generated, and therefore D(A)/S is well gener-
ated by Theorem 7.2.1.
Next observe that K(Q) induces an equivalence
∼
K(Mod A)/(Ker K(Q)) −→ K(A)
since K(Q) has K(T ) as a fully faithful right adjoint. Moreover, the cohomology
of each object in the kernel of K(Q) lies in the kernel of Q. Thus we obtain the
following commutative diagram.
Ker K(Q)
inc / K(Mod A) K(Q)
/ K(A)
can F

S
inc / D(A) can / D(A)/S
H∗ H̄

Ker Q
inc / Mod A Q
/A
It is easily checked that the kernel of F consists of all acyclic complexes. Thus
∼
F induces an equivalence D(A) − → D(A)/S.
7.8. Notes
Given a triangulated category T , there are two basic questions when one
studies exact localization functors T → T . One can ask for the existence of
a localization functor with some prescribed kernel, and one can ask for a
classification, or at least some structural results, for the set of all localization
functors on T . Well generated categories provide a suitable setting for some
partial answers.
The fact that cohomological functors induce localization functors is well
known for compactly generated triangulated categories [28], but the result
seems to be new for well generated categories. The localization theorem
which describes the localization with respect to a small set of objects is
due to Neeman [33]. The example of the derived category of an abelian
Grothendieck category is discussed in [3, 34]. The description of the ker-
nel of an exact functor between well generated categories seems to be new. A
motivation for this is the telescope conjecture which is due to Bousfield and
Ravenel [8, 40] and originally formulated for the stable homotopy category of
CW-spectra.
It is interesting to note that the existence of localization functors depends to
some extent on axioms from set theory; see for instance [11, 10].
224 Henning Krause
8. Epilogue: Beyond well generatedness

Well generated triangulated categories were introduced by Neeman as a
class of triangulated categories which includes all compactly generated cat-
egories and behaves well with respect to localization. We have discussed
in Sections 6 and 7 most of the basic properties of well generated cate-
gories but the picture is still not complete because some important ques-
tions remain open. For instance, given a well generated triangulated cate-
gory T , we do not know when a localizing subcategory arises as the kernel
of a localization functor and when it is generated by a small set of objects.
Also, one might ask when the set of all localizing subcategories is small.
Another aspect is Brown representability. We do know that every cohomolog-
ical functor T op → Ab preserving small products is representable, but what
about covariant functors T → Ab? It seems that one obtains more insight by
studying the universal cohomological functors T → Aα (T ); in particular we
need to know when they are full; see [43, 35] for some recent work in this
direction.
Instead of answering these open questions, let us be adventurous and move
a little bit beyond the class of well generated categories. In fact, there are
natural examples of triangulated categories which are not well generated. Such
examples arise from additive categories by taking their homotopy category of
chain complexes. More precisely, let A be an additive category and suppose
that A admits small coproducts. We denote by K(A) the category of chain
complexes in A whose morphisms are the homotopy classes of chain maps.
Take for instance the category A = Ab of abelian groups. Then one can show
that K(Ab) is not well generated; see [33]. In fact, more is true. The category
K(Ab) admits no small set of generators, that is, any localizing subcategory
generated by a small set of objects is a proper subcategory. However, it is
not difficult to show that any localizing subcategory generated by a small
set of objects is well generated. So we may think of K(Ab) as locally well
generated. In fact, discussions with Jan Šťovı́ček suggest that K(A) is locally
well generated whenever A is locally finitely presented; see [46]. Recall that
A is locally finitely presented if A has filtered colimits and there exists a small
set of finitely presented objects A0 such that every object can be written as the
filtered colimit of objects in A0 . On the other hand, K(A) is only generated
by a small set of objects if A = Add A0 for some small set of objects A0 .
Here, Add A0 denotes the smallest subcategory of A which contains A0 and is
closed under taking small coproducts and direct summands. We refer to [46]
for further details.
Appendix A. The abelianization of a triangulated category

Let C be an additive category. We consider functors F : C op → Ab into the
category of abelian groups and call a sequence F → F → F of functors
exact if the induced sequence F X → F X → F X of abelian groups is exact
for all X in C. A functor F is said to be coherent if there exists an exact sequence
(called presentation)
C(−, X) −→ C(−, Y ) −→ F −→ 0.
The morphisms between two coherent functors form a small set by Yoneda’s
lemma, and the coherent functors C op → Ab form an additive category with
cokernels. We denote this category by C.
A basic tool is the fully faithful Yoneda functor hC : C → C which sends
an object X to C(−, X). One might think of this functor as the completion of
C with respect to the formation of finite colimits. To formulate some further
properties, we recall that a morphism X → Y is a weak kernel for a morphism
Y → Z if the induced sequence C(−, X) → C(−, Y ) → C(−, Z) is exact.
Lemma A.1. Let C be an additive category.
(1) Given an additive functor H : C → A to an additive category which admits

cokernels, there is (up to a unique isomorphism) a unique right exact functor
H̄ : C → A such that H = H̄ ◦ hC .
(2) If C has weak kernels, then C is an abelian category.
(3) If C has small coproducts, then C has small coproducts and the Yoneda
functor preserves small coproducts.
Proof. (1) Extend H to H̄ by sending F in C with presentation

(−,φ)
C(−, X) −→ C(−, Y ) −→ F −→ 0
to the cokernel of H φ.
(2) The category C has cokernels, and it is therefore sufficient to show
that C has kernels. To this end fix a morphism F1 → F2 with the following
presentation.
C(−, X1 ) / C(−, Y1 ) / F1 /0

C(−, X2 ) / C(−, Y2 ) / F2 /0
226 Henning Krause
We construct the kernel F0 → F1 by specifying the following presentation.

C(−, X0 ) / C(−, Y0 ) / F0 /0

C(−, X1 ) / C(−, Y1 ) / F1 /0
First the morphism Y0 → Y1 is obtained from the weak kernel sequence

Y0 −→ X2 Y1 −→ Y2 .
Then the morphisms X0 → X1 and X0 → Y0 are obtained from the weak kernel
sequence
X0 −→ X1 Y0 −→ Y1 .
(3) For every family of functors Fi having a presentation
(−,φi )
C(−, Xi ) −→ C(−, Yi ) −→ Fi −→ 0,
%
the coproduct F = i Fi has a presentation
& (−,φi ) &
C(−, Xi ) −→ C(−, Yi ) −→ F −→ 0.
i i
Thus coproducts in C are not computed pointwise.

The assigment C
→ C is functorial in the following weak sense. Given a
functor F : C → D, there is (up to a unique isomorphism) a unique right exact
: C → D
functor F extending the composite hD ◦ F : C → D.
Now let T be a triangulated category. Then we write A(T ) = T and call
this category the abelianization of T , because the Yoneda functor T → A(T )
is the universal cohomological functor for T .
Lemma A.2. Let T be a triangulated category. Then the category A(T ) is
abelian and the Yoneda functor hT : T → A(T ) is cohomological.
(1) Given a cohomological functor H : T → A to an abelian category, there
is (up to a unique isomorphism) a unique exact functor H̄ : A(T ) → A
such that H = H̄ ◦ hT .
(2) Given an exact functor F : T → T between triangulated categories, there
is (up to a unique isomorphism) a unique exact functor A(F ) : A(T ) →
A(T ) such that hT ◦ F = A(F ) ◦ hT .
Proof. The category T has weak kernels and therefore A(T ) is abelian. Note
that the weak kernel of a morphism Y → Z is obtained by completing the
morphism to an exact triangle X → Y → Z → SX.
(1) Let H : T → A be a cohomological functor and let H̄ : A(T ) → A be

the right exact functor extending H which exists by Lemma A.1. Then H̄ is
exact because H is cohomological.
(2) Let F : T → T be exact. Then H = hT ◦ F is a cohomological functor
and we let A(F ) = H̄ be the exact functor which extends H .
The assignment T
→ A(T ) from triangulated categories to abelian cat-
egories preserves various properties of exact functors between triangulated
categories. Let us mention some of them.
Lemma A.3. Let F : T → T and G : T → T be exact functors between
triangulated categories.
(1) F is fully faithful if and only if A(F ) is fully faithful.
∼
(2) If F induces an equivalence T / Ker F − → T , then A(F ) induces an equiv-
∼
alence A(T )/(Ker A(F )) − → A(T ).
(3) F preserves small (co)products if and only if A(F ) preserves small
(co)products.
(4) F is left adjoint to G if and only if A(F ) is left adjoint to A(G).
Proof. Straightforward.
Notes
The abelianization of a triangulated category appears in Verdier’s thèse [48]
and in Freyd’s work on the stable homotopy category [15]. Note that their
construction is slightly different from the one given here, which is based on
coherent functors in the sense of Auslander [4].
Appendix B. Locally presentable abelian categories

Fix a regular cardinal α and a small additive category C which admits α-
coproducts. We denote by Add(C op , Ab) the category of additive functors
C op → Ab into the category of abelian groups. This is an abelian category
which admits small (co)products. In fact, (co)kernels and (co)products are
computed pointwise in Ab. Given functors F and G in Add(C op , Ab), we write
HomC (F, G) for the set of morphisms F → G. The most important objects in
Add(C op , Ab) are the representable functors C(−, X) with X ∈ C. Recall that
Yoneda’s lemma provides a bijection
∼
HomC (C(−, X), F ) −→ F X
for all F : C op → Ab and X ∈ C.
228 Henning Krause
We denote by Addα (C op , Ab) the full subcategory of Add(C op , Ab) which is

formed by all functors preserving α-products. This is an exact abelian subcate-
gory, because kernels and cokernels of morphism between α-product preserv-
ing functors preserve α-products. In particular, Addα (C op , Ab) is an abelian
category.
Now suppose that C admits cokernels. Then Lexα (C op , Ab) denotes the
full subcategory of Add(C op , Ab) which is formed by all left exact functors
preserving α-products. This category is locally presentable in the sense of
Gabriel and Ulmer and we refer to [17, §5] for an extensive treatment. In this
appendix we collect some basic facts.
First observe that α-filtered colimits in Lexα (C op , Ab) are computed point-
wise. This follows from the fact that in Ab taking α-filtered colimits commutes
with taking α-limits; see [17, Satz 5.12]. In particular, Lexα (C op , Ab) has small
coproducts because every small coproduct is the α-filtered colimit of its sub-
coproducts with less than α factors.
Next we show that one can identify Addα (C op , Ab) with a category of left
exact functors. To this end consider the Yoneda functor hC : C → C taking X
to C(−, X).
Lemma B.1. Let C be a small additive category with α-coproducts. Then the
Yoneda functor induces an equivalence
∼
Lexα (Cop , Ab) −→ Addα (C op , Ab)
by taking a functor F to F ◦ hC .
Proof. Use that every additive functor C op → Ab extends uniquely to a left

exact functor Cop → Ab; see Lemma A.1.
From now on we assume that C admits α-coproducts and cokernels.

Given any additive functor F : C op → Ab, we consider the category C/F
whose objects are pairs (C, μ) consisting of an object C ∈ C and an element
μ ∈ F C. A morphism (C, μ) → (C , μ ) is a morphism φ : C → C such that
F φ(μ ) = μ.
Lemma B.2. Let F : C op → Ab be an additive functor.
(1) The canonical morphism
colim C(−, C) −→ F
−−−→
(C,μ)∈C/F
in Add(C op , Ab) is an isomorphism.

(2) The functor F belongs to Lexα (C op , Ab) if and only if the category C/F is
α-filtered.
Proof. (1) is easy. For (2), see [17, Satz 5.3].
The representable functors in Lexα (C op , Ab) share the following finiteness

property. Recall that an object X from an additive category A with α-filtered
colimits is α-presentable if the representable functor A(X, −) : A → Ab pre-
serves α-filtered colimits. Next observe that the inclusion Lexα (C op , Ab) →
Add(C op , Ab) preserves α-filtered colimits. This follows from the fact that
in Ab taking α-filtered colimits commutes with taking α-limits. This has the
following consequence.
Lemma B.3. For each X in C, the representable functor C(−, X) is an α-

presentable object of Lexα (C op , Ab).
Proof. Combine Yoneda’s lemma with the fact that the inclusion
Lexα (C op , Ab) → Add(C op , Ab) preserves α-filtered colimits.
There is a general result for the category Lexα (C op , Ab) which says that tak-
ing α-filtered colimits commutes with taking α-limits; see [17, Korollar 7.12].
Here we need the following special case.
Lemma B.4. Suppose the category Lexα (C op , Ab) is abelian. Then an α-filtered
colimit of exact sequences is again exact.
Proof. We need to show that taking α-filtered colimits commutes with taking
kernels and cokernels. A cokernel is nothing but a colimit and therefore taking
colimits and cokernels commute. The statement about kernels follows from
the fact that the inclusion Lexα (C op , Ab) → Add(C op , Ab) preserves kernels
and α-filtered colimits. Thus we can compute kernels and α-filtered colimits
in Add(C op , Ab) and therefore in the category Ab of abelian groups. In Ab it is
well known that taking kernels and filtered colimits commute.
Lemma B.5. Suppose that C is abelian. Then Lexα (C op , Ab) is abelian and the
Yoneda functor hC : C → Lexα (C op , Ab) is exact. Given an abelian category
A which admits small coproducts and exact α-filtered colimits, and given a
functor F : Lexα (C op , Ab) → A preserving α-filtered colimits, we have that F
is exact if and only if F ◦ hC is exact.
Proof. We use the analogue of Lemma B.2 for morphisms which says that each
morphism φ in Lexα (C op , Ab) can be written as α-filtered colimit φ = colim φi
−−−→
i∈C/φ
230 Henning Krause
of morphisms between representable functors. Thus one computes
Coker φ = colim Coker φi and Ker φ = colim Ker φi ,

−−−→ −−−→
i∈C/φ i∈C/φ
and we see that Lexα (C op , Ab) is abelian; see Lemma B.4. The formula for
kernels and cokernels shows that each exact sequence can be written as α-
filtered colimit of exact sequences in the image of the Yoneda embedding. The
criterion for the exactness of a functor Lexα (C op , Ab) → A is an immediate
consequence.
Let A be a cocomplete additive category. We denote by Aα the full subcate-

gory which is formed by all α-presentable objects. Following [17], the category
A is called locally α-presentable if Aα is small and each object is an α-filtered
colimit of α-presentable objects. We call A locally presentable if it is locally β-
presentable for some cardinal β. Note that we have for each locally presentable

category A a filtration A = β Aβ where β runs through all regular cardinals.
We have already seen that Lexα (C op , Ab) is locally α-presentable, and the next
lemma implies that, up to an equivalence, all locally α-presentable categories
are of this form.
Let f : C → A be a fully faithful and right exact functor into a cocomplete
additive category. Suppose that f preserves α-coproducts and that each object
in the image of f is α-presentable. Then f induces the functor
f∗ : A −→ Lexα (C op , Ab), X
→ A(f −, X),
and the following lemma discusses its left adjoint.
Lemma B.6. There is a fully faithful functor f ∗ : Lexα (C op , Ab) → A which

sends each representable functor C(−, X) to f X and identifies Lexα (C op , Ab)
with the full subcategory of A formed by all colimits of objects in the image of
f . The functor f ∗ is a left adjoint of f∗ .
Proof. The functor is the left Kan extension of f ; it takes F = colim C(−, C)
−−−→
(C,μ)∈C/F
in Lexα (C op , Ab) to colim f C in A. We refer to [17, Satz 7.8] for details.
−−−→
(C,μ)∈C/F
Suppose now that C is a triangulated category. The following lemma char-

acterizes the cohomological functors C op → Ab.
Lemma B.7. Let C be a small triangulated category and suppose C

admits α-coproducts. For a functor F in Addα (C op , Ab) the following are
equivalent.
(1) The category C/F is α-filtered.

(2) F is an α-filtered colimit of representable functors.
(3) F is a cohomological functor.
Proof. The implications (1) ⇒ (2) ⇒ (3) are clear. So we prove (3) ⇒ (1). It
is convenient to identify Addα (C op , Ab) with Lexα (Cop , Ab) and this identifies
F with the left exact functor F̄ : C → Ab which extends F . In fact, F̄ is exact
since F is cohomological, by Lemma A.2. Now write F̄ as α-filtered colimit of
representable functors F̄ = colim C(−, M); see Lemma B.2. The exactness
−−−→
F̄
(M,ν)∈C/
of F̄ implies that the representable functors C(−, C) with C ∈ C form a full
F̄ which is cofinal. We identify this subcategory with C/F
subcategory of C/
and conclude from Lemma 6.4.1 that C/F is α-filtered.
Next we discuss the functoriality of the assignment C

→ Addα (C op , Ab).
Lemma B.8. Let f : C → D be an exact functor between small triangulated

categories which admit α-coproducts. Suppose that f preserves α-coproducts.
Then the restriction functor
f∗ : Addα (Dop , Ab) −→ Addα (C op , Ab), F

→ F ◦ f,
has a left adjoint f ∗ which sends C(−, X) to D(−, f X) for all X in C. Moreover,
the following holds.
(1) The functors f∗ and f ∗ are exact.

∼
(2) Suppose f induces an equivalence C/ Ker f − → D. Then f∗ is fully faithful.
(3) Suppose f is fully faithful. Then f ∗ is fully faithful. Moreover, a cohomo-
logical functor F : Dop → Ab is in the essential image of f ∗ if and only if
every morphism D(−, D) → F factors through D(−, f C) for some object
C in C.
(4) A cohomological functor F : C op → Ab belongs to the kernel of f ∗ if
and only if every morphism C(−, C) → F factors through a morphism
C(−, γ ) : C(−, C) → C(−, C ) such that f γ = 0.
Proof. The left adjoint of f∗ is the left Kan extension. We can describe it explic-
itly if we identify Addα (C op , Ab) with Lexα (Cop , Ab); see Lemma B.1. Given a
functor F in Lexα (Cop , Ab) written as α-filtered colimit F = colim C(−, C)
−−−→

(C,μ)∈C/F
of representable functors, we put

f ∗ F = colim D(−, fC).
−−−→

(C,μ)∈C/F
232 Henning Krause
Thus f ∗ makes the following diagram commutative.
hC =
C
hC
/ C / Lex (Cop , Ab) / Addα (C op , Ab)
α
f f f∗ f∗
hD
hD
/D / Lex (D = / Addα (Dop , Ab)
D α
op , Ab)
We check that f ∗ is a left adjoint for f∗ . For a representable functor F =

X) we have
C(−,
X), G) = HomD(D(−,
HomD(f ∗ C(−, fX), G) ∼
= G(fX)
= f∗ G(X) ∼ X), f∗ G)
= HomC(C(−,
for all G in Lexα (Dop , Ab). Clearly, this isomorphism extends to every colimit
of representable functors.
(1) The exactness of f∗ is clear because a sequence F → F → F in
Addα (C op , Ab) is exact if and only if F X → F X → F X is exact for all X in
C. For the exactness of f ∗ we identify again Add(C op , Ab) with Lexα (Cop , Ab)
and apply Lemma B.5. Thus we need to check that the composition of f ∗ with
the Yoneda functor hC is exact. But we have that f ∗ ◦ hC = hD ◦ f, and now the
exactness follows from that of f . Finally, we use the fact that taking α-filtered
colimits in Addα (Dop , Ab) is exact by Lemma B.4.
(2) It is well known that for any epimorphism f : C → D of addi-
tive categories inducing a bijection Ob C → Ob D, the restriction functor
Add(Dop , Ab) → Add(C op , Ab) is fully faithful; see [29, Corollary 5.2]. Given
a triangulated subcategory C ⊆ C, the quotient functor C → C/C is an epimor-
phism. Thus the assertion follows since Addα (C op , Ab) is a full subcategory of
Add(C op , Ab).
(3) We keep our identification Addα (C op , Ab) = Lexα (Cop , Ab) and consider
the adjunction morphism η : Id → f∗ ◦ f ∗ . We claim that η is an isomorphism.
Because f is fully faithful, ηF is an isomorphism for each representable functor
X). It follows that ηF is an isomorphism for all F since f ∗ and f∗
F = C(−,
both preserve α-filtered colimits and each F can be expressed as α-filtered
colimit of representable functors. Now Proposition 2.3.1 implies that f ∗ is
fully faithful.
Let F be a cohomological functor in Addα (Dop , Ab) and apply Lemma B.7 to
write the functor as α-filtered colimit F = colim D(−, D) of representable
−−−→
(D,μ)∈D/F
functors. Suppose first that every morphism D(−, D) → F factors through
D(−, f C) for some C ∈ C. Then Im f/F is a cofinal subcategory of D/F
and therefore F = colim D(−, D) by Lemma 6.4.1. Thus F belongs to

−−−→
(D,μ)∈Im f/F
the essential image of f since D(−, f C) = f ∗ C(−, C) for all C ∈ C and the
∗
essential image is closed under taking colimits. Now suppose that F belongs
to the essential image of f ∗ . Then F = f ∗ G ∼
= f ∗ f∗ f ∗ G = f ∗ f∗ F for some
G. The functor f∗ F is cohomological and therefore f∗ F = colim C(−, C),
−−−→
(C,μ)∈C/f∗ F
∼
again by Lemma B.7. Thus F = colim D(−, f C) and we use Lemma B.3
−−−→
(C,μ)∈C/f∗ F
to conclude that each morphism D(−, D) → F factors through D(−, f C) for
some (C, μ) ∈ C/f∗ F .
(4) Let F be a cohomological functor in Addα (C op , Ab) and apply
Lemma B.7 to write the functor as α-filtered colimit F = colim C(−, C)
−−−→
(C,μ)∈C/F
of representable functors. Now f ∗ F = colim D(−, f C) = 0 if and only
−−−→
(C,μ)∈C/F
if for each D ∈ D, we have colim D(D, f C) = 0. This happens iff for
−−−→
(C,μ)∈C/F
each (C, μ) ∈ C/F , we find a morphism γ : C → C in C/F inducing a map
D(f C, f C) → D(f C, f C ) which annihilates the identity morphism. But this
means that f γ = 0 and that μ : C(−, C) → F factors through C(−, γ ).
Notes
Locally presentable categories were introduced and studied by Gabriel and
Ulmer in [17]; see [1] for a modern treatment. In [33], Neeman initiated the use
of locally presentable abelian categories for studying triangulated categories.
References
[1] J. Adámek and J. Rosický, Locally presentable and accessible categories,
Cambridge Univ. Press, Cambridge, 1994.
[2] J. F. Adams, Stable homotopy and generalised homology, Univ. Chicago Press,
Chicago, Ill., 1974.
[3] L. Alonso Tarrı́o, A. Jeremı́as López and M. J. Souto Salorio, Localization in
categories of complexes and unbounded resolutions, Canad. J. Math. 52 (2000),
no. 2, 225–247.
[4] M. Auslander, Coherent functors, in Proc. Conf. Categorical Algebra (La Jolla,
Calif., 1965), 189–231, Springer, New York, 1966.
[5] L. Avramov and S. Halperin, Through the looking glass: a dictionary between
rational homotopy theory and local algebra, in Algebra, algebraic topology and
their interactions (Stockholm, 1983), 1–27, Lecture Notes in Math., 1183, Springer,
Berlin, 1986.
234 Henning Krause
[6] A. A. Beı̆linson, J. Bernstein and P. Deligne, Faisceaux pervers, in Analysis and

topology on singular spaces, I (Luminy, 1981), 5–171, Astérisque, 100, Soc. Math.
France, Paris, 1982.
[7] D. Benson, S. B. Iyengar and H. Krause, Local cohomology and support for
triangulated categories, Ann. Sci. École Norm. Sup. (4) 41 (2008), no. 4, 575–
621.
[8] A. K. Bousfield, The localization of spectra with respect to homology, Topology
18 (1979), no. 4, 257–281.
[9] E. H. Brown, Jr., Cohomology theories, Ann. of Math. (2) 75 (1962), 467–484.
[10] C. Casacuberta, J. J. Gutiérrez and J. Rosický, Are all localizing subcategories of
stable homotopy categories coreflective?, preprint (2007).
[11] C. Casacuberta, D. Scevenels and J. H. Smith, Implications of large-cardinal
principles in homotopical localization, Adv. Math. 197 (2005), no. 1, 120–139.
[12] W. G. Dwyer, Localizations, in Axiomatic, enriched and motivic homotopy theory,
3–28, Kluwer Acad. Publ., Dordrecht, 2004.
[13] J. Franke, On the Brown representability theorem for triangulated categories,
Topology 40 (2001), no. 4, 667–680.
[14] P. Freyd, Abelian categories. An introduction to the theory of functors, Harper &
Row, New York, 1964.
[15] P. Freyd, Stable homotopy, in Proc. Conf. Categorical Algebra (La Jolla, Calif.,
1965), 121–172, Springer, New York, 1966.
[16] P. Gabriel, Des catégories abéliennes, Bull. Soc. Math. France 90 (1962), 323–448.
[17] P. Gabriel and F. Ulmer, Lokal präsentierbare Kategorien, Lecture Notes in Math.,
221, Springer, Berlin, 1971.
[18] P. Gabriel and M. Zisman, Calculus of fractions and homotopy theory, Springer-
Verlag New York, Inc., New York, 1967.
[19] A. Grothendieck and J. L. Verdier, Préfaisceaux, in SGA 4, Théorie des Topos et
Cohomologie Etale des Schémas, Tome 1. Thórie des Topos, Lect. Notes in Math.,
vol. 269, Springer, Heidelberg, 1972-1973, pp. 1–184.
[20] A. Heider, Two results from Morita theory of stable model categories,
arXiv:0707.0707v1.
[21] B. Keller, Deriving DG categories, Ann. Sci. École Norm. Sup. (4) 27 (1994),
no. 1, 63–102.
[22] B. Keller, A remark on the generalized smashing conjecture, Manuscripta Math.
84 (1994), no. 2, 193–198.
[23] H. Krause, On Neeman’s well generated triangulated categories, Doc. Math. 6
(2001), 121–126 (electronic).
[24] H. Krause, A Brown representability theorem via coherent functors, Topology 41
(2002), no. 4, 853–861.
[25] H. Krause, Cohomological quotients and smashing localizations, Amer. J. Math.
127 (2005), no. 6, 1191–1246.
[26] A. Lévy, Basic set theory, Springer, Berlin, 1979.
[27] S. MacLane, Categories for the working mathematician, Springer, New York,
1971.
[28] H. R. Margolis, Spectra and the Steenrod algebra, North-Holland, Amsterdam,
1983.
[29] B. Mitchell, The dominion of Isbell, Trans. Amer. Math. Soc. 167 (1972), 319–331.
[30] A. Neeman, The connection between the K-theory localization theorem of Thoma-
son, Trobaugh and Yao and the smashing subcategories of Bousfield and Ravenel,
Ann. Sci. École Norm. Sup. (4) 25 (1992), no. 5, 547–566.
[31] A. Neeman, The Grothendieck duality theorem via Bousfield’s techniques and
Brown representability, J. Amer. Math. Soc. 9 (1996), no. 1, 205–236.
[32] A. Neeman, Non-compactly generated categories, Topology 37 (1998), no. 5,
981–987.
[33] A. Neeman, Triangulated categories, Ann. of Math. Stud., 148, Princeton Univ.
Press, Princeton, NJ, 2001.
[34] A. Neeman, On the derived category of sheaves on a manifold, Doc. Math. 6
(2001), 483–488 (electronic).
[35] A. Neeman, Brown representability follows from Rosický, preprint (2007).
[36] O. Ore, Linear equations in non-commutative fields, Ann. of Math. (2) 32 (1931),
no. 3, 463–477.
[37] M. Porta, The Popescu-Gabriel theorem for triangulated categories,
arXiv:0706.4458v1.
[38] D. Puppe, On the structure of stable homotopy theory, in Colloquium on algebraic
topology, Aarhus Universitet Matematisk Institut (1962), 65–71.
[39] D. G. Quillen, Homotopical algebra, Lecture Notes in Math., 43, Springer, Berlin,
1967.
[40] D. C. Ravenel, Localization with respect to certain periodic homology theories,
Amer. J. Math. 106 (1984), no. 2, 351–414.
[41] J. Rickard, Idempotent modules in the stable category, J. London Math. Soc. (2)
56 (1997), no. 1, 149–170.
[42] J. Rickard, Bousfield localization for representation theorists, in Infinite length
modules (Bielefeld, 1998), 273–283, Birkhäuser, Basel.
[43] J. Rosický, Generalized Brown representability in homotopy categories, Theory
Appl. Categ. 14 (2005), no. 19, 451–479 (electronic).
[44] N. Spaltenstein, Resolutions of unbounded complexes, Compositio Math. 65
(1988), no. 2, 121–154.
[45] B. Stenström, Rings of quotients, Springer, New York, 1975.
[46] J. Šťovı́ček, Locally well generated homotopy categories of complexes,
arXiv:0810.5684v1.
[47] G. Tabuada, Homotopy theory of well-generated algebraic triangulated categories,
J. K-theory, in press.
[48] J.-L. Verdier, Des catégories dérivées des catégories abéliennes, Astérisque No.
239 (1996), xii+253 pp. (1997).
Henning Krause, Institut für Mathematik, Universität Paderborn, 33095 Pader-

born, Germany.
E-mail address:
Homological algebra in bivariant K-theory
and other triangulated categories. I
ralf meyer and ryszard nest
Abstract. Bivariant (equivariant) K-theory is the standard setting for non-
commutative topology. We may carry over various techniques from homotopy
theory and homological algebra to this setting. Here we do this for some basic
notions from homological algebra: phantom maps, exact chain complexes, pro-
jective resolutions, and derived functors. We introduce these notions and apply
them to examples from bivariant K-theory.
An important observation of Beligiannis is that we can approximate our cat-
egory by an Abelian category in a canonical way, such that our homological
concepts reduce to the corresponding ones in this Abelian category. We compute
this Abelian approximation in several interesting examples, where it turns out to
be very concrete and tractable.
The derived functors comprise the second page of a spectral sequence that,
in favourable cases, converges towards Kasparov groups and other interest-
ing objects. This mechanism is the common basis for many different spectral
sequences. Here we only discuss the very simple 1-dimensional case, where the
spectral sequences reduce to short exact sequences.
1. Introduction
It is well-known that many basic constructions from homotopy theory extend to
categories of C∗ -algebras. As we argued in [17], the framework of triangulated
categories is ideal for this purpose. The notion of triangulated category was
introduced by Jean-Louis Verdier to formalise the properties of the derived
category of an Abelian category. Stable homotopy theory provides further clas-
sical examples of triangulated categories. The triangulated category structure
encodes basic information about manipulations with long exact sequences and
(total) derived functors. The main point of [17] is that the domain of the Baum–
Connes assembly map is the total left derived functor of the functor that maps
a G-C∗ -algebra A to K∗ (G r A).
2000 Mathematics Subject Classification. 18E30, 19K35, 46L80, 55U35.

This research was supported by the EU-Network Quantum Spaces and Noncommutative
Geometry (Contract HPRN-CT-2002-00280).
236
Homological algebra in bivariant K-theory 237
The relevant triangulated categories in non-commutative topology come

from Kasparov’s bivariant K-theory. This bivariant version of K-theory carries
a composition product that turns it into a category. The formal properties of
this and related categories are surveyed in [15], with an audience of homotopy
theorists in mind.
Projective resolutions are among the most fundamental concepts in homo-
logical algebra; several others like derived functors are based on it. Projective
resolutions seem to live in the underlying Abelian category and not in its derived
category. This is why total derived functor make more sense in triangulated
categories than the derived functors themselves. Nevertheless, we can define
derived functors in triangulated categories and far more general categories. This
goes back to Samuel Eilenberg and John C. Moore ([9]). We learned about this
theory in articles by Apostolos Beligiannis ([4]) and J. Daniel Christensen ([8]).
Homological algebra in non-Abelian categories is always relative, that is,
we need additional structure to get started. This is useful because we may fit
the additional data to our needs. In a triangulated category T, there are several
kinds of additional data that yield equivalent theories; following [8], we use an
ideal in T. We only consider ideals I in T of the form
I(A, B) := {x ∈ T(A, B) | F (x) = 0}
for a stable homological functor F : T → C into a stable Abelian category C.
Here stable means that C carries a suspension automorphism and that F inter-
twines the suspension automorphisms on T and C, and homological means that
exact triangles yield exact sequences. Ideals of this form are called homological
ideals.
A basic example is the ideal in the Kasparov category KK defined by
IK (A, B) := {f ∈ KK(A, B) | 0 = K∗ (f ) : K∗ (A) → K∗ (B)}. (1.1)
For a compact (quantum) group G, we define two ideals I ⊆ I,K ⊆ KKG
in the equivariant Kasparov category KKG by
I (A, B) := {f ∈ KKG (A, B) | G f = 0 in KK(G A, G B)},
(1.2)
I,K (A, B) := {f ∈ KKG (A, B) | K∗ (G f ) = 0}, (1.3)
where K∗ (G f ) denotes the map K∗ (G A) → K∗ (G B) induced by f .
For a locally compact group G and a (suitable) family of subgroups F, we
define the homological ideal
VC F (A, B) := {f ∈ KKG (A, B) | ResH
G (f ) = 0 in KK (A, B) for all H ∈ F}.
H
(1.4)
238 Ralf Meyer and Ryszard Nest
If F is the family of compact subgroups, then VC F is related to the Baum–

Connes assembly map ([17]). Of course, there are analogous ideals in more
classical categories of (spectra of) G-CW-complexes.
All these examples can be analysed using the machinery we explain. We
carry this out in some cases in Sections 4 and 5.
We use an ideal I to carry over various notions from homological algebra
to our triangulated category T. In order to see what they mean in examples, we
characterise them using a stable homological functor F : T → C with ker F =
I. This is often easy. For instance, a chain complex with entries in T is I-exact
if and only if F maps it to an exact chain complex in the Abelian category C
(see Lemma 28), and a morphism in T is an I-epimorphism if and only if F
maps it to an epimorphism. Here we may take any functor F with ker F = I.
But the most crucial notions like projective objects and resolutions require
a more careful choice of the functor F . Here we need the universal I-exact
functor, which is a stable homological functor F with ker F = I such that any
other such functor factors uniquely through F (up to natural equivalence). The
universal I-exact functor and its applications are due to Apostolos Beligiannis
([4]).
If F : T → C is universal, then F detects I-projective objects, and it iden-
tifies I-derived functors with derived functors in the Abelian category C (see
Theorem 59). Thus all our homological notions reduce to their counterparts in
the Abelian category C.
In order to apply this, we need to know when a functor F with ker F = I
is the universal one. We develop a new, useful criterion for this purpose here,
which uses partially defined adjoint functors (Theorem 57).
Our criterion shows that the universal IK -exact functor for the ideal IK
in KK in (1.1) is the K-theory functor K∗ , considered as a functor from KK to
the category AbZ/2
c of countable Z/2-graded Abelian groups (see Theorem 63).
Hence the derived functors for IK only involve Ext and Tor for Abelian groups.
For the ideal I,K in KKG in (1.3), we get the functor
KKG → Mod(Rep G)Z/2
c , A
→ K∗ (G A), (1.5)
Z/2
where Mod(Rep G)c denotes the Abelian category of countable Z/2-graded
modules over the representation ring Rep G of the compact (quantum) group G
(see Theorem 72); here we use a certain canonical Rep G-module structure on
K∗ (G A). Hence derived functors with respect to I,K involve Ext and Tor
for Rep G-modules.
We do not need the Rep G-module structure on K∗ (G A) to define I,K :
our machinery notices automatically that such a module structure is missing.
The universality of the functor in (1.5) clarifies in what sense homological
algebra with Rep G-modules is a linearisation of algebraic topology with G-C∗ -

algebras.
The universal homological functor for the ideal I is quite similar to the one
for I,K (see Theorem 73). There is a canonical Rep G-module structure on
G A as an object of KK, and the universal I -exact functor is essentially the
functor A
→ G A, viewed as an object of a suitable Abelian category that
encodes this Rep G-module structure; it also involves a fully faithful embedding
of KK in an Abelian category due to Peter Freyd ([10]).
The derived functors that we have discussed above appear in a spectral
sequence which – in favourable cases – computes morphism spaces in T (like
KKG (A, B)) and other homological functors. This spectral sequence is a gen-
eralisation of the Adams spectral sequence in stable homotopy theory and
is the main motivation for [8]. Much earlier, such spectral sequences were
studied by Hans-Berndt Brinkmann in [7]. In [16], this spectral sequence is
applied to our bivariant K-theory examples. Here we only consider the much
easier case where this spectral sequence degenerates to an exact sequence
(see Theorem 66). This generalises the familiar Universal Coefficient Theorem
for KK∗ (A, B).
2. Homological ideals in triangulated categories

After fixing some basic notation, we introduce several interesting ideals in
bivariant Kasparov categories; we are going to discuss these ideals throughout
this article. Then we briefly recall what a triangulated category is and introduce
homological ideals. Before we begin, we should point out that the choice of
ideal is important because all our homological notions depend on it. It seems
to be a matter of experimentation and experience to find the right ideal for a
given purpose.
2.1. Generalities about ideals in additive categories

All categories we consider will be additive, that is, they have a zero object and
finite direct products and coproducts which agree, and the morphism spaces
carry Abelian group structures such that the composition is additive in each
variable ([13]).
Notation 1. Let C be an additive category. We write C(A, B) for the group

of morphisms A → B in C, and A ∈∈ C to denote that A is an object of the
category C.
Definition 2. An ideal I in C is a family of subgroups I(A, B) ⊆ C(A, B) for

all A, B ∈∈ C such that
C(C, D) ◦ I(B, C) ◦ C(A, B) ⊆ I(A, D) for all A, B, C, D ∈∈ C.
We write I1 ⊆ I2 if I1 (A, B) ⊆ I2 (A, B) for all A, B ∈∈ C. Clearly, the

ideals in T form a complete lattice. The largest ideal C consists of all morphisms
in C; the smallest ideal 0 contains only zero morphisms.
Definition 3. Let C and C be additive categories and let F : C → C be an

additive functor. Its kernel ker F is the ideal in C defined by
ker F (A, B) := {f ∈ C(A, B) | F (f ) = 0}.
This should be distinguished from the kernel on objects, consisting of all

objects with F (A) ∼= 0, which is used much more frequently. The kernel is the
class of ker F -contractible objects that we introduce below.
Definition 4. Let I ⊆ T be an ideal. Its quotient category C/I has the same
objects as C and morphism groups C(A, B)/I(A, B).
The quotient category is again additive, and the obvious functor F : C →

C/I is additive and satisfies ker F = I. Thus any ideal I in C is of the form ker F
for a canonical additive functor F .
The additivity of C/I and F depends on the fact that any ideal I is compatible
with finite products in the following sense: the natural isomorphisms
∼
=
C(A, B1 × B2 ) −
→ C(A, B1 ) × C(A, B2 ),
∼
=
C(A1 × A2 , B) −
→ C(A1 , B) × C(A2 , B)
restrict to isomorphisms
∼
=
I(A, B1 × B2 ) −
→ I(A, B1 ) × I(A, B2 ),
∼
=
I(A1 × A2 , B) −
→ I(A1 , B) × I(A2 , B).
2.2. Examples of ideals

Example 5. Let KK be the Kasparov category, whose objects are the separable
C∗ -algebras and whose morphism spaces are the Kasparov groups KK0 (A, B),
with the Kasparov product as composition. Let AbZ/2 be the category of
Z/2-graded Abelian groups. Both categories are evidently additive.
K-theory is an additive functor K∗ : KK → AbZ/2 . We let IK := ker K∗ (as

in (1.1)). Thus IK (A, B) ⊆ KK(A, B) is the kernel of the natural map

γ : KK(A, B) → Hom K∗ (A), K∗ (B) := Hom Kn (A), Kn (B) .
n∈Z/2
There is another interesting ideal in KK, namely, the kernel of a natural

map

κ : IK (A, B) → Ext K∗ (A), K∗+1 (B) := Ext Kn (A), Kn+1 (B)
n∈Z/2
due to Lawrence Brown (see [23]),

whose definition we now recall.
We represent f ∈ KK(A, B) ∼
= Ext A, C0 (R, B) by a C∗ -algebra extension
C0 (R, B) ⊗ K E A. This yields an exact sequence
K1 (B) / K0 (E) / K0 (A)

O
f∗ f∗ (2.1)

K1 (A) o K1 (E) o K0 (B).
The vertical maps in (2.1) are the two components of γ (f ). If f ∈ IK (A, B),
then (2.1) splits into
two extensions of Abelian groups, which yield an ele-
ment κ(f ) in Ext K∗ (A), K∗+1 (B) .
Example 6. Let G be a second countable, locally compact group. Let KKG

be the associated equivariant Kasparov category; its objects are the separable
G-C∗ -algebras and its morphism spaces are the groups KKG (A, B), with the
Kasparov product as composition. If H ⊆ G is a closed subgroup, then there
is a restriction functor ResHG : KK → KK , which simply forgets part of the
G H
equivariance.
If F is a set of closed subgroups of G, we define an ideal VC F in KKG by
VC F (A, B) := {f ∈ KKG (A, B) | ResH

G (f ) = 0 for all H ∈ F}
as in (1.4). Of course, the condition ResH G (f ) = 0 is supposed to hold

in KKH (A, B). We are mainly interested in the case where F is the family
of all compact subgroups of G and simply denote the ideal by VC in this
case.
This ideal arises if we try to compute G-equivariant homology theories in
terms of H -equivariant homology theories for H ∈ F. The ideal VC is closely
related to the approach to the Baum–Connes assembly map in [17].
The authors feel more at home with Kasparov theory than with spectra. Many
readers will prefer to work in categories of spectra of, say, G-CW-complexes.
We do not introduce these categories here; but it shoud be clear enough that
they support similar restriction functors, which provide analogues of the ideals
VC F .
Example 7. Let G and KKG be as in Example 6. Using the crossed product

functor (also called descent functor)
G : KKG → KK, A
→ G A,
we define ideals I ⊆ I,K ⊆ KKG as in (1.2) and (1.3) by
I (A, B) := {f ∈ KKG (A, B) | G f = 0 in KK(G A, G B)},

I,K (A, B) := {f ∈ KKG (A, B) | K∗ (Gf ) = 0 : K∗ (GA) → K∗ (GB)}.
We only study these ideals for compact G. In this case, the Green–Julg Theo-
rem identifies K∗ (G A) with the G-equivariant K-theory KG ∗ (A) (see [11]).
Hence the ideal I,K ⊆ KKG is a good equivariant analogue of the ideal IK in
KK.
Literally the same definition as above provides ideals I ⊆ I,K ⊆ KKG
if G is a compact quantum group. We will always allow this more gen-
eral situation below, but readers unfamiliar with quantum groups may ignore
this.
Remark 8. We emphasise quantum groups here because Examples 6 and 7

become closely related in this context. This requires a quantum group ana-
logue of the ideals VC F in KKG of Example 6. If G is a locally com-
pact quantum group, then Saad Baaj and Georges Skandalis construct a
G-equivariant Kasparov category KKG in [2]. There is a forgetful functor
ResHG : KK → KK for each closed quantum subgroup H ⊆ G. Therefore,
G H
a family F of closed quantum subgroups yields an ideal VC F in KKG as in

Example 6.
Let G be a compact group as in Example 7. Any crossed product G A
carries a canonical coaction of G, that is, a coaction of the discrete quantum
group C∗ (G). Baaj–Skandalis duality asserts that this yields an equivalence of
categories KKG ∼
∗
= KKC (G) (see [2]). We get back the crossed product func-
tor KKG → KK by composing this equivalence with the restriction functor
∗
KKC (G) → KK for the trivial quantum subgroup. Hence I ⊆ KKG corre-
∗
sponds by Baaj–Skandalis duality to VC F ⊆ KKC (G) , where F consists only
of the trivial quantum subgroup.
Thus the constructions in Examples 6 and 7 are both special cases of a more
general construction for locally compact quantum groups.
Finally, we consider a classical example from homological algebra.

Example 9. Let C be an Abelian category. Let Ho(A) be the homotopy category
of unbounded chain complexes
δn δn−1 δn−2
· · · → Cn −
→ Cn−1 −−→ Cn−2 −−→ Cn−3 → · · ·
over C. The space of morphisms A → B in Ho(C) is the space [A, B] of
homotopy classes of chain maps from A to B.
Taking homology defines functors Hn : Ho(C) → C for n ∈ Z, which we
combine to a single functor H∗ : Ho(C) → CZ . We let IH ⊆ Ho(C) be its kernel:
IH (A, B) := {f ∈ [A, B] | H∗ (f ) = 0}. (2.2)
We also consider the category Ho(C; Z/p) of p-periodic chain complexes
over C for p ∈ N≥1 ; its objects satisfy Cn = Cn+p and δn = δn+p for all n ∈ Z,
and chain maps and homotopies are required p-periodic as well. The cate-
gory Ho(C; Z/2) plays a role in connection with cyclic cohomology, especially
with local cyclic cohomology ([14, 21]). The category Ho(C; Z/1) is isomor-
phic to the category of chain complexes without grading. By convention, we
let Z/0 = Z, so that Ho(C; Z/0) = Ho(C).
The homology of a periodic chain complex is, of course, periodic, so that
we get a homological functor H∗ : Ho(C; Z/p) → CZ/p ; here CZ/p denotes the
category of Z/p-graded objects of C. We let IH ⊆ Ho(C; Z/p) be the kernel
of H∗ as in (2.2).
2.3. What is a triangulated category?

A triangulated category is a category T with a suspension automorphism
: T → T and a class of exact triangles, subject to various axioms (see [17,
19, 25]). An exact triangle is a diagram in T of the form
A → B → C → A or A \9 /B
9
[1]99

C,
where the [1] in the arrow C → A warns us that this map has degree 1. A
morphism of triangles is a triple of maps α, β, γ making the obvious diagram
commute.
A typical example is the homotopy category Ho(C; Z/p)) of Z/p-graded
chain complexes. Here the suspension functor is the (signed) translation functor

(Cn , dn ) := (Cn−1 , −dn−1 ) on objects,

(fn ) := (fn−1 ) on morphisms;
a triangle is exact if it is isomorphic to a mapping cone triangle

f
A → B → cone(f ) → A
for some chain map f ; the maps B → cone(f ) → A are the canonical ones. It
is well-known that this defines a triangulated category for p = 0; the arguments
for p ≥ 1 are essentially the same.
Another classical example is the stable homotopy category, say, of compactly
generated pointed topological spaces (it is not particularly relevant which cate-
gory of spaces or spectra we use). The suspension is (A) := S1 ∧ A; a triangle
is exact if it is isomorphic to a mapping cone triangle
f
A → B → cone(f ) → A
for some map f ; the maps B → cone(f ) → A are the canonical ones.
We are mainly interested in the categories KK and KKG introduced in §2.2.
Their triangulated category structure is discussed in detail in [17]. We are facing
a notational problem because the functor X
→ C0 (X) from pointed compact
spaces to C∗ -algebras is contravariant, so that mapping cone triangles now
have the form
f
A ← B ← cone(f ) ← C0 (R, A)
for a ∗-homomorphism f : B → A; here

, -
cone(f ) = (a, b) ∈ C0 (0, ∞], A × B a(∞) = f (b)
and the maps C0 (R, A) → cone(f ) → B are the obvious ones, a

→ (a, 0) and
(a, b)
→ b.
It is reasonable to view a ∗-homomorphism from A to B as a morphism
from B to A. Nevertheless, we prefer the convention that an algebra homor-
phism A → B is a morphism A → B. But then the most natural triangulated
category structure lives on the opposite category KKop . This creates only nota-
tional difficulties because the opposite category of a triangulated category
inherits a canonical triangulated category structure, which has “the same”
exact triangles. However, the passage to opposite categories exchanges suspen-
sions and desuspensions and modifies some sign conventions. Thus the functor
A
→ C0 (R, A), which is the suspension functor in KKop , becomes the desus-
pension functor in KK. Fortunately, Bott periodicity implies that 2 ∼ = id, so
−1
that and agree.
Depending on your definition of a triangulated category, you may want the
suspension to be an equivalence or isomorphism of categories. In the latter
case, you must replace KK(G) by an equivalent category (see [17]); since this
is not important here, we do not bother about this issue.
A triangle in KK(G) is called exact if it is isomorphic to a mapping cone

triangle
f
C0 (R, B) → cone(f ) → A → B
for some (equivariant) ∗-homomorphism f .
An important source of exact triangles in KKG are extensions. If A B
C is an extension of G-C∗ -algebras with an equivariant completely positive
contractive section, then it yields a class in Ext(C, A) ∼
= KK( −1 C, A); the
resulting triangle
−1 C → A → B → C
in KKG is exact and called an extension triangle. It is easy to see that any exact
triangle is isomorphic to an extension triangle.
It is shown in [17] that KK and KKG for a locally compact group G are
triangulated categories with this extra structure. The same holds for the equiv-
ariant Kasparov theory KKS with respect to any C∗ -bialgebra S; this theory
was defined by Baaj and Skandalis in [2].
The triangulated category axioms are discussed in greater detail in [17,
19, 25]. They encode some standard machinery for manipulating long exact
sequences. Most of them amount to formal properties of mapping cones and
mapping cylinders, which we can prove as in classical topology. The only
axiom that requires more care is that any morphism f : A → B should be part
of an exact triangle.
Unlike in [17], we prefer to construct this triangle as an extension trian-
gle because this works in greater generality; we have taken this idea from
Radu Popescu and Alexander Bonkat ([6, 20]). Any element in KKS0 (A, B) ∼ =
KKS1 A, C0 (R, B) can be represented by an extension K(H) E A with
an equivariant completely positive contractive section, where H is a full
S-equivariant Hilbert C0 (R, B)-module, so that K(H) is KKS -equivalent to
C0 (R, B). Hence the resulting extension triangle in KKS is isomorphic to one
of the form
C0 (R, A) → C0 (R, B) → E → A;
by construction, it contains the suspension of the given class in KKS0 (A, B); it
is easy to remove the suspension.
Definition 10. Let T be a triangulated and C an Abelian category. A covariant
functor F : T → C is called homological if F (A) → F (B) → F (C) is exact
at F (B) for all exact triangles A → B → C → A. A contravariant functor
with the analogous exactness property is called cohomological.
Let A → B → C → A be an exact triangle. Then a homological functor

F : T → C yields a natural long exact sequence
· · · → Fn+1 (C) → Fn (A) → Fn (B) → Fn (C) → Fn−1 (A) → Fn−1 (B) → · · ·
with Fn (A) := F ( −n A) for n ∈ Z, and a cohomological functor F : Top → C

yields a natural long exact sequence
· · · ← F n+1 (C) ← F n (A) ← F n (B) ← F n (C) ← F n−1 (A) ← F n−1 (B) ← · · ·
with F n (A) := F ( −n A).
Proposition 11. Let T be a triangulated category. The functors
T(A, ) : T → Ab, B
→ T(A, B)
are homological for all A ∈∈ T. Dually, the functors
T( , B) : Top → Ab, A
→ T(A, B)
are cohomological for all B ∈∈ T.
Observe that
Tn (A, B) = T( −n A, B) ∼
= T(A, n B) ∼
= T−n (A, B).
Definition 12. A stable additive category is an additive category equipped with

an (additive) automorphism , called suspension.
A stable homological functor is a homological functor F : T → C into a
stable Abelian category C together with natural isomorphisms F (A) ∼
=
T
C F (A) for all A ∈∈ T.
Example 13. The category CZ/p of Z/p-graded objects of an Abelian cate-

gory C is stable for any p ∈ N; the suspension automorphism merely shifts the
grading. The functors K∗ : KK → AbZ/2 and H∗ : Ho(C; Z/p) → CZ/p intro-
duced in Examples 5 and 9 are stable homological functors.
If F : T → C is any homological functor, then

F∗ : T → CZ , A
→ Fn (A) n∈Z
is a stable homological functor. Many of our examples satisfy Bott period-

icity, that is, there is a natural isomorphism F2 (A) ∼
= F (A). Then we get a
stable homological functor F∗ : T → CZ/2 . A typical example for this is the
functor K∗ .
Definition 14. A functor F : T → T between two triangulated categories is

called exact if it intertwines the suspension automorphisms (up to specified
natural isomorphisms) and maps exact triangles in T again to exact triangles
in T .
Example 15. The restriction functor ResHG : KK → KK for a closed quan-

G H
tum subgroup H of a locally compact quantum group G and the crossed product
functors G , G r : KKG → KK are exact because they preserve mapping
cone triangles.
Let F : T1 → T2 be an exact functor. If G : T2 →? is exact, homological,

or cohomological, then so is G ◦ F .
Using Examples 13 and 15, we see that the functors that define the ideals
ker γ in Example 5, VC F in Example 6, I , and I,K in Example 7, and IH in
Example 9 are all stable and either homological or exact.
2.4. The universal homological functor

The following general construction of Peter Freyd ([10]) plays an important
role in [4]. For an additive category C, let Fun(Cop , Ab) be the category of
contravariant additive functors C → Ab, with natural transformations as mor-
phisms. Unless C is essentially small, this is not quite a category because the
morphisms may form classes instead of sets. We may ignore this set-theoretic
problem because the bivariant Kasparov categories that we are interested in
are essentially small, and the subcategory Coh(C) of Fun(Cop , Ab) that we are
going to use later on is an honest category for any C.
The category Fun(Cop , Ab) is Abelian: if f : F1 → F2 is a natural transfor-
mation, then its kernel, cokernel, image, and co-image are computed pointwise
on the objects of C, so that they boil down to the corresponding constructions
with Abelian groups.
The Yoneda embedding is an additive functor
Y : C → Fun(Cop , Ab), B
→ T( , B).
This functor is fully faithful, and there are natural isomorphisms
Hom(Y(B), F ) ∼
= F (B) for all F ∈∈ Fun(Cop , Ab), B ∈∈ T
by the Yoneda Lemma. A functor F ∈∈ Fun(Cop , Ab) is called representable

if it is isomorphic to Y(B) for some B ∈∈ C. Hence Y yields an equiva-
lence of categories between C and the subcategory of representable functors in
Fun(Cop , Ab).
A functor F ∈∈ Fun(Cop , Ab) is called finitely presented if there is an exact

sequence Y(B1 ) → Y(B2 ) → F → 0 with B1 , B2 ∈∈ T. Since Y is fully faith-
ful, this means that F is the cokernel of Y(f ) for a morphism f in C. We let
Coh(C) be the full subcategory of finitely presented functors in Fun(Cop , Ab).
Since representable functors belong to Coh(C), we still have a Yoneda embed-
ding Y : C → Coh(C). Although the category Coh(T) tends to be very big and
therefore unwieldy, it plays an important theoretical role.
Theorem 16 (Freyd’s Theorem). Let T be a triangulated category.

Then Coh(T) is a stable Abelian category that has enough projective and
enough injective objects, and the projective and injective objects coincide.
The functor Y : T → Coh(T) is fully faithful, stable, and homological. Its
essential range Y(T) consists of projective-injective objects. Conversely, an
object of Coh(T) is projective-injective if and only if it is a retract of an object
of Y(T).
The functor Y is the universal (stable) homological functor in the fol-
lowing sense: any (stable) homological functor F : T → C to a (stable)
Abelian category C factors uniquely as F = F̄ ◦ Y for a (stable) exact functor
F : Coh(T) → C .
If idempotents in T split – as in all our examples – then Y(T) is closed

under retracts, so that Y(T) is equal to the class of projective-injective objects
in Coh(T).
2.5. Homological ideals in triangulated categories

Let T be a triangulated category, let C be a stable additive category, and let
F : T → C be a stable homological functor. Then ker F is a stable ideal in the
following sense:
Definition 17. An ideal I in T is called stable if the suspension isomorphisms

∼
=
: T(A, B) −
→ T(A, B) for A, B ∈∈ T restrict to isomorphisms
∼
=
: I(A, B) −
→ I(A, B).
If I is stable, then there is a unique suspension automorphism on T/I for

which the canonical functor T → T/I is stable. Thus the stable ideals are
exactly the kernels of stable additive functors.
Definition 18. An ideal I in a triangulated category T is called homological if

it is the kernel of a stable homological functor.
Remark 19. Freyd’s Theorem shows that Y induces a bijection between (sta-
ble) exact functors Coh(T) → C and (stable) homological functors T → C
because F̄ ◦ Y is homological if F̄ : Coh(T) → C is exact. Hence the notion
of homological functor is independent of the triangulated category structure
on T because the Yoneda embedding Y : T → Coh(T) does not involve any
additional structure. Hence the notion of homological ideal only uses the sus-
pension automorphism, not the class of exact triangles.
All the ideals considered in §2.2 except for ker κ in Example 5 are ker-
nels of stable homological functors or exact functors. Those of the first kind are
homological by definition. If F : T → T is an exact functor between two trian-
gulated categories, then Y ◦ F : T → Coh(T ) is a stable homological functor
with ker Y ◦ F = ker F by Freyd’s Theorem 16. Hence kernels of exact func-
tors are homological as well.
Is any homological ideal the kernel of an exact functor? This is not the
case:
Proposition 20. Let Der(Ab) be the derived category of the category Ab of
Abelian groups. Define the ideal IH in Der(Ab) as in Example 9. This ideal is
not the kernel of an exact functor.
We postpone the proof to the end of §3.1 because it uses the machinery
of §3.1.
It takes some effort to characterise homological ideals because T/I is almost
never Abelian. The results in [4, §2–3] show that an ideal is homological if
and only if it is saturated in the notation of [4]. We do not discuss this notion
here because most ideals that we consider are obviously homological. The
only example where we could profit from an abstract characterisation is the
ideal ker κ in Example 5.
There is no obvious homological functor whose kernel is ker κ because κ is
not a functor on KK. Nevertheless, ker κ is the kernel of an exact functor; the
relevant functor is the functor KK → UCT, where UCT is the variant of KK
that satisfies the Universal Coefficient Theorem in complete generality. This
functor can be constructed as a localisation of KK (see [17]). The Universal
Coefficient Theorem implies that its kernel is exactly ker κ.
3. From homological ideals to derived functors

Once we have a stable homological functor F : T → C, it is not surprising
that we can do a certain amount of homological algebra in T. For instance,
we may call a chain complex of objects of T F -exact if F maps it to an exact
chain complex in C; and we may call an object F -projective if F maps it to a

projective object in C. But are these definitions reasonable?
We propose that a reasonable homological notion should depend only on the
ideal ker F . We will see that the notion of F -exact chain complex is reasonable
and only depends on ker F . In contrast, the notion of projectivity above depends
on F and is only reasonable in special cases. There is another, more useful,
notion of projective object that depends only on the ideal ker F .
Various notions from homological algebra still make sense in the context of
homological ideals in triangulated categories. Our discussion mostly follows
[1, 4, 8, 9]. All our definitions involve only the ideal, not a stable homological
functor that defines it. We reformulate them in terms of an exact or a stable
homological functor defining the ideal in order to understand what they mean
in concrete cases. Following [9], we construct projective objects using adjoint
functors.
The most sophisticated concept in this section is the universal I-exact func-
tor, which gives us complete control over projective resolutions and derived
functors. We can usually describe such functors very concretely.
3.1. Basic notions

We introduce some useful terminology related to an ideal:
Definition 21. Let I be a homological ideal in a triangulated category T.

f g
– Let f : A → B be a morphism in T; embed it in an exact triangle A → B →
h
C → A. We call f
∗ I-monic if h ∈ I;
∗ I-epic if g ∈ I;
∗ an I-equivalence if it is both I-monic and I-epic, that is, g, h ∈ I;
∗ an I-phantom map if f ∈ I.
– An object A ∈∈ T is called I-contractible if idA ∈ I(A, A).
f g h
– An exact triangle A → B → C → A in T is called I-exact if h ∈ I.
The notions of monomorphism (or monic morphism) and epimorphism (or

epic morphism) – which can be found in any book on category theory such
as [13] – are categorical ways to express injectivity or surjectivity of maps. A
morphism in an Abelian category that is both monic and epic is invertible.
The classes of I-phantom maps, I-monics, I-epics, and of I-exact triangles
determine each other uniquely because we can embed any morphism in an exact
triangle in any position. It is a matter of taste which of these is considered most
fundamental. Following Daniel Christensen ([8]), we favour the phantom maps.

Other authors prefer exact triangles instead ([1, 4, 9]). Of course, the notion of
an I-phantom map is redundant; it becomes more relevant if we consider, say,
the class of I-exact triangles as our basic notion.
Notice that f is I-epic or I-monic if and only if −f is. If f is I-epic or
I-monic, then so are n (f ) for all n ∈ Z because I is stable. Similarly, (signed)
suspensions of I-exact triangles remain I-exact triangles.
Lemma 22. Let F : T → C be a stable homological functor into a stable

Abelian category C.
– A morphism f in T is
∗ a ker F -phantom map if and only if F (f ) = 0;
∗ ker F -monic if and only if F (f ) is monic;
∗ ker F -epic if and only if F (f ) is epic;
∗ a ker F -equivalence if and only if F (f ) is invertible.
– An object A ∈∈ T is ker F -contractible if and only if F (A) = 0.
– An exact triangle A → B → C → A is ker F -exact if and only if
0 → F (A) → F (B) → F (C) → 0
is a short exact sequence in C.

0 f f 0
Proof. Sequences in C of the form X → Y → Z or X → Y → Z are exact
at Y if and only if f is monic or epic, respectively. Moreover, a sequence of the
0 0
form X → Y → Z → U → W is exact if and only if 0 → Y → Z → U → 0
is exact.
Combined with the long exact homology sequences for F and suitable
exact triangles, these observations yield the assertions about monomorphisms,
epimorphisms, and exact triangles. The description of equivalences and con-
tractible objects follows, and phantom maps are trivial, anyway.
Now we specialise these notions to the ideal IK ⊆ KK of Example 5, replac-

ing IK by K in our notation to avoid clutter.
– Let f ∈ KK(A, B) and let K∗ (f ) : K∗ (A) → K∗ (B) be the induced map.

Then f is
∗ a K-phantom map if and only if K∗ (f ) = 0;
∗ K-monic if and only if K∗ (f ) is injective;
∗ K-epic if and only if K∗ (f ) is surjective;
∗ a K-equivalence if and only if K∗ (f ) is invertible.
– A C∗ -algebra A ∈∈ KK is K-contractible if and only if K∗ (A) = 0.

– An exact triangle A → B → C → A in KK is K-exact if and only if
0 → K∗ (A) → K∗ (B) → K∗ (C) → 0
is a short exact sequence (of Z/2-graded Abelian groups).
Similar things happen for the other ideals in §2.2 that are naturally defined
as kernels of stable homological functors.
Remark 23. It is crucial for the above theory that we consider functors that
are both stable and homological. Everything fails if we drop
either assumption
and consider functors such as K0 (A) or Hom Z/4, K∗ (A) .
Lemma 24. An object A ∈∈ T is I-contractible if and only if 0 : 0 → A is

an I-equivalence. A morphism f in T is an I-equivalence if and only if its
generalised mapping cone is I-contractible.
Thus the classes of I-equivalences and of I-contractible objects determine

each other. But they do not allow us to recover the ideal itself. For instance,
the ideals IK and ker κ in Example 5 have the same contractible objects and
equivalences.
Proof. Recall that the generalised mapping cone of f is the object C that fits
f
in an exact triangle A → B → C → A. The long exact sequence for this
triangle yields that F (f ) is invertible if and only if F (C) = 0, where F is
some stable homological functor F with ker F = I. Now the second assertion
follows from Lemma 22. Since the generalised mapping cone of 0 → A is A,
the first assertion is a special case of the second one.
Many ideals are defined as ker F for an exact functor F : T → T between

triangulated categories. We can also use such a functor to describe the above
notions:
Lemma 25. Let T and T be triangulated categories and let F : T → T be an

exact functor.
– A morphism f ∈ T(A, B) is
∗ a ker F -phantom map if and only if F (f ) = 0;
∗ ker F -monic if and only if F (f ) is (split) monic.
∗ ker F -epic if and only if F (f ) is (split) epic;
∗ a ker F -equivalence if and only if F (f ) is invertible.
– An object A ∈∈ T is ker F -contractible if and only if F (A) = 0.
– An exact triangle A → B → C → A is ker F -exact if and only if the exact
triangle F (A) → F (B) → F (C) → F (A) in T splits.
We will explain the notation during the proof.
Proof. A morphism f : X → Y in T is called split epic (split monic) if there

f g
is g : Y → X with f ◦ g = idY (g ◦ f = idX ). An exact triangle X → Y →
h
Z → X is said to split if h = 0. This immediately yields the characterisation
of ker F -exact triangles. Any split triangle is isomorphic to a direct sum triangle,
so that f is split monic and g is split epic ([19, Corollary 1.2.7]). Conversely,
either of these conditions implies that the triangle is split.
Since the ker F -exact triangles determine the ker F -epimorphisms and
ker F -monomorphisms, the latter are detected by F (f ) being split epic or
split monic, respectively. It is clear that split epimorphisms and split mono-
morphisms are epimorphisms and monomorphisms, respectively. The converse
holds in a triangulated category because if we embed a monomorphism or
epimorphism in an exact triangle, then one of the maps is forced to vanish, so
that the exact triangle splits.
Finally, a morphism is invertible if and only if it is both split monic and split
epic, and the zero map F (A) → F (A) is invertible if and only if F (A) = 0.
Alternatively, we may prove Lemma 22 using the Yoneda embedding

Y : T → Coh(T ). The assertions about phantom maps, equivalences, and con-
tractibility boil down to the observation that Y is fully faithful. The assertions
about monomorphisms and epimorphisms follow because a map f : A → B
in T becomes epic (monic) in Coh(T ) if and only if it is split epic (monic)
in T .
'
There is a similar description for ker Fi for a set {Fi } of exact functors.
This applies to the ideal VC F for a family of (quantum) subgroups F in a
locally compact (quantum) group G (Example 6). Replacing VC F by F in our
notation to avoid clutter, we get:
– A morphism f ∈ KKG (A, B) is

∗ an F-phantom map if and only if ResH G (f ) = 0 in KK for all H ∈ F;
H
G (f ) is (split) epic in KK for all H ∈ F;

∗ F-epic if and only if ResH H
∗ F-monic if and only if ResG (f ) is (split) monic in KKH for all H ∈ F;

H
∗ an F-equivalence if and only if ResH H

G (f ) is a KK -equivalence for all
H ∈ F.
– A G-C∗ -algebra A ∈∈ KKG is F-contractible if and only if ResH ∼
G (A) = 0
in KK for all H ∈ F.
H
– An exact triangle A → B → C → A in KKG is F-exact if and only if
G (A) → ResG (B) → ResG (C) → ResG (A)

ResH H H H
is a split exact triangle in KKH for all H ∈ F.

You may write down a similar list for the ideal I ⊆ KKG of
Example 7.
Lemma 25 allows us to prove that the ideal IH in Der(Ab) cannot be the
kernel of an exact functor:
Proof of Proposition 20. We embed Ab → Der(Ab) as chain complexes

concentrated in degree 0. The generator τ ∈ Ext(Z/2, Z/2) corresponds to
the extension of Abelian groups Z/2 Z/4 Z/2, where the first map is
multiplication by 2 and the second map is the natural projection. We get an
exact triangle
τ
Z/2 → Z/4 → Z/2 −
→ Z/2[1]
in Der(Ab). This triangle is IH -exact because the map Z/2 → Z/4 is injective
as a group homomorphism and hence IH -monic in Der(Ab).
Assume there were an exact functor F : Der(Ab) → T with ker F = IH .
Then F (τ ) = 0, so that F maps our triangle to a split triangle and F (Z/4) ∼ =
F (Z/2) ⊕ F (Z/2) by Lemma 25. It follows that F (2 · idZ/4 ) = 2 · idF (Z/4) = 0
because 2 · idF (Z/2) = F (2 · idZ/2 ) = 0. Hence 2 · idZ/4 ∈ ker F = IH , which
is false. This contradiction shows that there is no exact functor F with ker F =
IH .
One of the most interesting questions about an ideal is whether all

I-contractible objects vanish or, equivalently, whether all I-equivalences are
invertible. These two questions are equivalent by Lemma 24. The answer is
negative for the ideal IK in KK because the Universal Coefficient Theorem
does not hold for arbitrary separable C∗ -algebras. Therefore, we also get coun-
terexamples for the ideal I,K in KKG for a compact quantum group. In
contrast, if G is a connected Lie group with torsion-free fundamental group,
then I -equivalences in KKG are invertible (see [18]). If G is an amenable
group, then VC-equivalences in KKG are invertible; this follows from the proof
of the Baum–Connes Conjecture for these groups by Nigel Higson and Gennadi
Kasparov (see [17]). These examples show that this question is subtle and may
involve difficult analysis.
3.2. Exact chain complexes

The notion of I-exactness, which we have only defined for exact triangles
so far, will now be extended to chain complexes. Our definition differs from
Beligiannis’ one ([1, 4]), which we recall first.
Let T be a triangulated category and let I be a homological ideal in T.
Definition 26. A chain complex
dn+1 dn dn−1
C• := (· · · → Cn+1 −−→ Cn −
→ Cn−1 −−→ Cn−2 → · · · )
in T is called I-decomposable if there is a sequence of I-exact triangles
gn fn hn
Kn+1 −
→ Cn −
→ Kn −
→ Kn+1
with dn = gn−1 ◦ fn : Cn → Cn−1 .
Such complexes are called I-exact in [1, 4]. This definition is inspired by the
following well-known fact: a chain complex over an Abelian category is exact
if and only if it splits into short exact sequences of the form Kn Cn Kn−1
as in Definition 26.
We prefer another definition of exactness because we have not found a
general explicit criterion for a chain complex to be I-decomposable.
Definition 27. Let C• = (Cn , dn ) be a chain complex over T. For each n ∈ N,

embed dn in an exact triangle
dn fn gn
Cn −
→ Cn−1 −
→ Xn −
→ Cn . (3.1)
gn fn+1
We call C• I-exact in degree n if the map Xn − → Cn −−−→ Xn+1 belongs
to I(Xn , Xn+1 ). This does not depend on auxiliary choices because the exact
triangles in (3.1) are unique up to (non-canonical) isomorphism.
We call C• I-exact if it is I-exact in degree n for all n ∈ Z.
This definition is designed to make the following lemma true:
Lemma 28. Let F : T → C be a stable homological functor into a stable

Abelian category C with ker F = I. A chain complex C• over T is I-exact in
degree n if and only if
F (dn+1 ) F (dn )
F (Cn+1 ) −−−−→ F (Cn ) −−−→ F (Cn−1 )
is exact at F (Cn ).
Proof. The complex C• is I-exact in degree n if and only if the map
−1 F (gn ) F (fn+1 )
−1 F (Xn ) −−−−−→ F (Cn ) −−−−→ F (Xn+1 )
vanishes. Equivalently, the range of −1 F (gn ) is contained in the kernel of

F (fn+1 ). The long exact sequences
−1 F (gn ) F (dn )
· · · → −1 F (Xn ) −−−−−→ F (Cn ) −−−→ F (Cn−1 ) → · · · ,
F (dn+1 ) F (fn+1 )
· · · → F (Cn+1 ) −−−−→ F (Cn ) −−−−→ F (Xn+1 ) → · · ·
show that the range of −1 F (gn ) and the kernel of F (fn+1 ) are equal to the
kernel of F (dn ) and the range of F (dn+1 ), respectively. Hence C• is I-exact in
degree n if and only if ker F (dn ) ⊆ range F (dn+1 ). Since dn ◦ dn+1 = 0, this is
equivalent to ker F (dn ) = range F (dn+1 ).
Corollary 29. I-decomposable chain complexes are I-exact.
Proof. Let F : T → C be a stable homological functor with ker F = I. If C•

is I-decomposable, then F (C• ) is obtained by splicing short exact sequences
in C. This implies that F (C• ) is exact, so that C• is I-exact by Lemma 28.
The converse implication in Corollary 29 fails in general (see Example 37).
Example 30. For the ideal IK in KK, Lemma 28 yields that a chain complex C•
over KK is K-exact (in degree n) if and only if the chain complex
· · · → K∗ (Cn+1 ) → K∗ (Cn ) → K∗ (Cn−1 ) → · · ·
of Z/2-graded Abelian groups is exact (in degree n). Similar remarks apply
to the other ideals in §2.2 that are defined as kernels of stable homological
functors.
As a trivial example, we consider the largest possible ideal I = T. This
ideal is defined by the zero functor. Lemma 28 or the definition yield that
all chain complexes are T-exact. In contrast, it seems hard to characterise the
I-decomposable chain complexes, already for I = T.
Lemma 31. A chain complex of length 3

f g
··· → 0 → A −
→B−
→ C → 0 → ···
f g
is I-exact if and only if there are an I-exact exact triangle A −
→ B −
→ C →
A and a commuting diagram
f g
A / B / C
∼ α ∼ β ∼ γ (3.2)

A
f
/B g
/C
where the vertical maps α, β, γ are I-equivalences. Furthermore, we can

achieve that α and β are identity maps.
Proof. Let F be a stable homological functor with I = ker F .

Suppose first that we are in the situation of (3.2). Lemma 22 yields that F (α),
F (β), and F (γ ) are invertible and that 0 → F (A ) → F (B ) → F (C ) → 0 is
a short exact sequence. Hence so is 0 → F (A) → F (B) → F (C) → 0. Now
Lemma 28 yields that our given chain complex is I-exact.
Conversely, suppose that we have an I-exact chain complex. By Lemma 28,
this means that 0 → F (A) → F (B) → F (C) → 0 is a short exact sequence.
Hence f : A → B is I-monic. Embed f in an exact triangle A → B → C →
A. Since f is I-monic, this triangle is I-exact. Let α = idA and β = idB .
Since the functor T( , C) is cohomological and g ◦ f = 0, we can find a map
γ : C → C making (3.2) commute. The functor F maps the rows of (3.2) to
short exact sequences by Lemmas 28 and 22. Now the Five Lemma yields that
F (γ ) is invertible, so that γ is an I-equivalence.
Remark 32. Lemma 31 implies that I-exact chain complexes of length 3 are
I-decomposable. We do not expect this for chain complexes of length 4. But
we have not searched for a counterexample.
Which chain complexes over T are I-exact for I = 0 and hence for any
homological ideal? The next definition provides the answer.
Definition 33. A chain complex C• over a triangulated category is called

homologically exact if F (C• ) is exact for any homological functor F : T → C.
Example 34. If A → B → C → A is an exact triangle, then the chain

complex
· · · → −1 A → −1 B → −1 C → A → B → C → A → B → C → · · ·
is homologically exact by the definition of a homological functor.
Lemma 35. Let F : T → T be an exact functor between two triangulated

categories. Let C• be a chain complex over T. The following are equivalent:
(1) C• is ker F -exact in degree n;

(2) F (C• ) is I-exact in degree n with respect to I = 0;
(3) the chain complex Y ◦ F (C• ) in Coh(T ) is exact in degree n;
(4) F (C• ) is homologically exact in degree n;
(5) the chain complexes of Abelian groups T A, F (C• ) are exact in degree n
for all A ∈∈ T .
Proof. By Freyd’s Theorem 16, Y ◦ F : T → Coh(T ) is a stable homological

functor with ker F = ker(Y ◦ F ). Hence Lemma 28 yields (1) ⇐⇒ (3). Simi-
larly, we have (2) ⇐⇒ (3) because Y : T → Coh(T ) is a stable homological
functor with ker Y = 0. Freyd’s Theorem 16 also asserts that any homological
functor F : T → C factors as F̄ ◦ Y for an exact functor F̄ . Hence (3)=⇒(4).
Proposition 11 yields (4)=⇒(5). Finally, (5) ⇐⇒ (3) because kernels and cok-
ernels in Coh(T ) are computed pointwise on objects of T .
Remark 36. More generally, consider a set of exact functors Fi : T → Ti . As

in the proof of the equivalence (1) ⇐⇒ (2) in Lemma 35, we see that a chain
'
complex C• is ker Fi -exact (in degree n) if and only if the chain complexes
Fi (C• ) are exact (in degree n) for all i.
As a consequence, a chain complex C• over KKG for a locally compact

quantum group G is F-exact if and only if ResH G (C• ) is homologically exact
for all H ∈ F. A chain complex C• over KKG for a compact quantum group G
is I -exact if and only if the chain complex G C• over KK is homologically
exact.
Example 37. We exhibit an I-exact chain complex that is not I-decomposable

for the ideal I = 0. By Lemma 25, any 0-exact triangle is split. Therefore,
a chain complex is 0-decomposable if and only if it is a direct sum of chain
id
complexes of the form 0 → Kn − → Kn → 0. Hence any decomposable chain
complex is contractible and therefore mapped by any homological functor to a
contractible chain complex. By the way, if idempotents in T split then a chain
complex is 0-decomposable if and only if it is contractible.
As we have remarked in Example 34, the chain complex
· · · → −1 C → A → B → C → A → B → C → 2 A → · · ·
is homologically exact for any exact triangle A → B → C → A. But such

chain complexes need not be contractible. A counterexample is the exact trian-
gle Z/2 → Z/4 → Z/2 → Z/2 in Der(Ab), which we have already used in
the proof of Proposition 20. The resulting chain complex over Der(Ab) cannot
be contractible because H∗ maps it to a non-contractible chain complex.
3.2.1. More homological algebra with chain complexes

Using our notion of exactness for chain complexes, we can do homological
algebra in the homotopy category Ho(T). We briefly sketch some results in
this direction, assuming some familiarity with more advanced notions from
homological algebra. We will not use this later.
The I-exact chain complexes form a thick subcategory of Ho(T) because

of Lemma 28. We let Der := Der(T, I) be the localisation of Ho(T) at this
subcategory and call it the derived category of T with respect to I.
We let Der≥n and Der≤n be the full subcategories of Der consisting of
chain complexes that are I-exact in degrees less than n and greater than n,
respectively.
Theorem 38. The pair of subcategories Der≥0 , Der≤0 forms a truncation

structure (t-structure) on Der in the sense of [3].
Proof. The main issue here is the truncation of chain complexes. Let C• be a
chain complex over T. We embed the map d0 in an exact triangle C0 → C−1 →
X → C0 and let C•≥0 be the chain complex
· · · → C2 → C1 → C0 → C−1 → X → C0 → C−1 → X

→ 2 C0 → · · · .
This chain complex is I-exact – even homologically exact – in negative degrees,

that is, C•≥0 ∈ Der≥0 . The triangulated category structure allows us to construct
a chain map C•≥0 → C• that is an isomorphism on Cn for n ≥ −1. Hence its
mapping cone C•≤−1 is I-exact – even contractible – in degrees ≥ 0, that is,
C•≤−1 ∈∈ Der≤−1 . By construction, we have an exact triangle
C•≥0 → C• → C•≤−1 → C•≥0
in Der.
We also have to check that there is no non-zero morphism C• → D• in
Der if C• ∈∈ Der≥0 and D• ∈∈ Der≤−1 . Recall that morphisms in Der are
∼
represented by diagrams C• ← C̃• → D• in Ho(T), where the first map is
an I-equivalence. Hence C̃• ∈∈ Der≥0 as well. We claim that any chain map
f : C̃•≥0 → D•≤−1 is homotopic to 0. Since the maps C̃•≥0 → C• and D• →
D•≤−1 are I-equivalences, any morphism C• → D• vanishes in Der.
It remains to prove the claim. In a first step, we use that D•≤−1 is contractible
in degrees ≥ 0 to replace f by a homotopic chain map supported in degrees <
0. In a second step, we use that C̃•≥0 is homologically exact in the relevant
degrees to recursively construct a chain homotopy between f and 0.
Any truncation structure gives rise to an Abelian category, its core. The core
of Der is the full subcategory C of all chain complexes that are I-exact except
in degree 0. This is a stable Abelian category, and the standard embedding
T → Ho(T) yields a stable homological functor F : T → C with ker F = I.
This functor is characterised uniquely by the following universal property:

any (stable) homological functor H : T → C with I ⊆ ker H factors uniquely
as H = H̄ ◦ F for an exact functor H̄ : C → C . We construct H̄ in three steps.
First, we lift H to an exact functor Ho(H ) : Ho(T, I) → Ho(C ). Secondly,
Ho(H ) descends to a functor Der(H ) : Der(T, I) → Der(C ). Finally, Der(H )
restricts to a functor H̄ : C → C between the cores. Since I ⊆ ker H , an
I-exact chain complex is also ker H -exact. Hence Ho(H ) preserves exact-
ness of chain complexes by Lemma 28. This allows us to construct Der(H )
and shows that Der(H ) is compatible with truncation structures. This allows
us to restrict it to an exact functor between the cores. Finally, we use that the
core of the standard truncation structure on Der(C) is C. It is easy to see that
we have H̄ ◦ F = H .
Especially, we get an exact functor Der(F ) : Der(T, I) → Der(C), which
restricts to the identity functor idC between the cores. Hence Der(F ) is fully
faithful on the thick subcategory generated by C ⊆ Der(T, I). It seems plau-
sible that Der(F ) should be an equivalence of categories under some mild
conditions on I and T.
We will continue our study of the functor F : T → C in §3.7. The uni-
versal property determines it uniquely. Beligiannis ([4]) has another, simpler
construction.
3.3. Projective objects

Let I be a homological ideal in a triangulated category T.
Definition 39. A homological functor F : T → C is called I-exact if F (f ) = 0

for all I-phantom maps f or, equivalently, I ⊆ ker F . An object A ∈∈ T is
called I-projective if the functor T(A, ) : T → Ab is I-exact. Dually, an object
B ∈∈ T is called I-injective if the functor T( , B) : T → Abop is I-exact.
We write PI for the class of I-projective objects in T.
The notions of projective and injective object are dual to each other: if
we pass to the opposite category Top with the canonical triangulated category
structure and use the same ideal Iop , then this exchanges the roles of projective
and injective objects. Therefore, it suffices to discuss one of these two notions
in the following. We will only treat projective objects because all the ideals
in §2.2 have enough projective objects, but most of them do not have enough
injective objects.
Notice that the functor F is I-exact if and only if the associated stable
functor F∗ : T → CZ is I-exact because I is stable.
Since we require F to be homological, the long exact homology sequence

and Lemma 28 yield that the following conditions are all equivalent to F being
I-exact:
– F maps I-epimorphisms to epimorphisms in C;

– F maps I-monomorphisms to monomorphisms in C;
– 0 → F (A) → F (B) → F (C) → 0 is a short exact sequence in C for any
I-exact triangle A → B → C → A;
– F maps I-exact chain complexes to exact chain complexes in C.
This specialises to equivalent definitions of I-projective objects.
Lemma 40. An object A ∈∈ T is I-projective if and only if I(A, B) = 0 for

all B ∈∈ T.
Proof. If f ∈ I(A, B), then f = f∗ (idA ). This has to vanish if A is I-projective.

Suppose, conversely, that I(A, B) = 0 for all B ∈∈ T. If f ∈ I(B, B ), then
T(A, f ) maps T(A, B) to I(A, B ) = 0, so that T(A, f ) = 0. Hence A is
I-projective.
An I-exact functor also has the following properties (which are strictly
weaker than being I-exact):
– F maps I-equivalences to isomorphisms in C;

– F maps I-contractible objects to 0 in C.
Again we may specialise this to I-projective objects.
Lemma 41. The class of I-exact homological functors T → Ab or T → Abop

is closed under composition with ±1 : T → T, retracts, direct sums, and direct
products. The class PI of I-projective objects is closed under (de)suspensions,
retracts, and possibly infinite direct sums (as far as they exist in T).
Proof. The first assertion follows because direct sums and products of Abelian
groups are exact; the second one is a special case.
Notation 42. Let P be a set of objects of T. We let (P)⊕ be the smallest class
of objects of T that contains P and is closed under retracts and direct sums (as
far as they exist in T).
By Lemma 41, (P)⊕ consists of I-projective objects if P does. We say

that P generates all I-projective objects if (P)⊕ = PI . In examples, it is
usually easier to describe a class of generators in this sense.
3.4. Projective resolutions

Definition 43. Let I ⊆ T be a homological ideal in a triangulated category and
let A ∈∈ T. A one-step I-projective resolution is an I-epimorphism π : P → A
with P ∈∈ PI . An I-projective resolution of A is an I-exact chain complex
δn+1 δn δn−1 δ1 δ0
· · · −−→ Pn −
→ Pn−1 −−→ · · · −
→ P0 −
→A
with Pn ∈∈ PI for all n ∈ N.

We say that I has enough projective objects if each A ∈∈ T has a one-step
I-projective resolution.
The following proposition contains the basic properties of projective reso-

lutions, which are familiar from the similar situation for Abelian categories.
Proposition 44. If I has enough projective objects, then any object of T has
an I-projective resolution (and vice versa).
Let P• → A and P• → A be I-projective resolutions. Then any map A →
A may be lifted to a chain map P• → P• , and this lifting is unique up to chain

homotopy. Two I-projective resolutions of the same object are chain homotopy
equivalent. As a result, the construction of projective resolutions provides a
functor
P : T → Ho(T).
f g h
Let A → B → C → A be an I-exact triangle. Then there exists a canon-
ical map η : P (C) → P (A)[1] in Ho(T) such that the triangle
P (f ) P (g) η
P (A) −−→ P (B) −−→ P (C) −
→ P (A)[1]
in Ho(T) is exact; here [1] denotes the translation functor in Ho(T), which has
nothing to do with the suspension in T.
Proof. Let A ∈∈ T. By assumption, there is a one-step I-projective resolution

δ0 : P0 → A, which we embed in an exact triangle A1 → P0 → A → A1 .
Since δ0 is I-epic, this triangle is I-exact. By induction, we construct a
sequence of such I-exact triangles An+1 → Pn → An → An+1 for n ∈ N
with Pn ∈∈ P and A0 = A. By composition, we obtain maps δn : Pn → Pn−1
for n ≥ 1, which satisfy δn ◦ δn+1 = 0 for all n ≥ 0. The resulting chain com-
plex
δn δn−1 δ1 δ0
· · · → Pn −
→ Pn−1 −−→ Pn−2 → · · · → P1 −
→ P0 −
→A→0
is I-decomposable by construction and therefore I-exact by Corollary 29.

The remaining assertions are proved exactly as their classical counterparts

in homological algebra. We briefly sketch the arguments. Let P• → A and
P• → A be I-projective resolutions and let f ∈ T(A, A ). We construct fn ∈
T(Pn , Pn ) by induction on n such that the diagrams
P0 @
δ0
/A Pn C
δn
/ Pn−1
@@ CC
@@ CC
f0 @@ f and fn CC fn−1 for n ≥ 1
@ C!
P0 / A Pn / Pn−1

δ0

δn
commute. We must check that this is possible. Since the chain complex P• → A
is I-exact and Pn is I-projective for all n ≥ 0, the chain complexes

(δm )∗ (δ0 )∗
· · · → T(Pn , Pm ) −−→ T(Pn , Pm−1

) → · · · → T(Pn , P0 ) −−→ T(Pn , A) → 0
are exact for all n ∈ N. This allows us to find maps fn as above. By con-
struction, these maps form a chain map lifting f : A → A . Its uniqueness up
to chain homotopy is proved similarly. If we apply this unique lifting result
to two I-projective resolutions of the same object, we get the uniqueness of
I-projective resolutions up to chain homotopy equivalence. Hence we get a
well-defined functor P : T → Ho(T).
Now consider an I-exact triangle A → B → C → A as in the third
paragraph of the lemma. Let X• be the mapping cone of some chain map
P (A) → P (B) in the homotopy class P (f ). This chain complex is supported
in degrees ≥ 0 and has I-projective entries because Xn = P (A)n−1 ⊕ P (B)n .
The map X0 = 0 ⊕ P (B)0 → B → C yields a chain map X• → C, that is, the
composite map X1 → X0 → C vanishes. By construction, this chain map lifts
the given map B → C and we have an exact triangle P (A) → P (B) → X• →
P (A)[1] in Ho(T). It remains to observe that X• → C is I-exact. Then X• is
an I-projective resolution of C. Since such resolutions are unique up to chain
homotopy equivalence, we get a canonical isomorphism X• ∼ = P (C) in Ho(T)
and hence the assertion in the third paragraph.
Let F be a stable homological functor with I = ker F . We have to check that
F (X• ) → F (C) is a resolution. This
reduces
to a well-known
diagram
chase in
Abelian categories, using that F P (A) → F (A) and F P (B) → F (B) are
resolutions and that F (A) F (B) F (C) is exact.
3.5. Derived functors

We only define derived functors if there are enough projective objects because
this case is rather easy and suffices for our applications. The general case can
be reduced to the familiar case of Abelian categories using the results of §3.2.1.
Definition 45. Let I be a homological ideal in a triangulated category T

with enough projective objects. Let F : T → C be an additive functor with
values in an Abelian category C. It induces a functor Ho(F ) : Ho(T) → Ho(C),
applying F pointwise to chain complexes. Let P : T → Ho(T) be the projective
resolution functor constructed in Proposition 44. Let Hn : Ho(C) → C be the
nth homology functor for some n ∈ N. The composite functor
P Ho(F ) Hn
Ln F : T −
→ Ho(T) −−−→ Ho(C) −→ C
is called the nth left derived functor of F . If F : Top → C is an additive functor,
then the corresponding functor H n ◦ Ho(F ) ◦ P : Top → C is denoted by Rn F
and called the nth right derived functor of F .
More concretely, let A ∈∈ T and let (P• , δ• ) be an I-projective resolution
of A. If F is covariant, then Ln F (A) is the homology at F (Pn ) of the chain
complex
F (δn+1 ) F (δn )
· · · → F (Pn+1 ) −−−−→ F (Pn ) −−→ F (Pn−1 ) → · · · → F (P0 ) → 0.
If F is contravariant, then Rn F (A) is the cohomology at F (Pn ) of the cochain
complex
F (δn+1 ) F (δn )
· · · ← F (Pn+1 ) ←−−−− F (Pn ) ←−− F (Pn−1 ) ← · · · ← F (P0 ) ← 0.
Lemma 46. Let A → B → C → A be an I-exact triangle. If F : T → C is
a covariant additive functor, then there is a long exact sequence
· · · → Ln F (A) → Ln F (B) → Ln F (C) → Ln−1 F (A)

→ · · · → L1 F (C) → L0 F (A) → L0 F (B) → L0 F (C) → 0.
If F is contravariant instead, then there is a long exact sequence
· · · ← Rn F (A) ← Rn F (B) ← Rn F (C) ← Rn−1 F (A)

← · · · ← R1 F (C) ← R0 F (A) ← R0 F (B) ← R0 F (C) ← 0.
Proof. This follows from the third assertion of Proposition 44 together with the
well-known long exact homology sequence for exact triangles in Ho(C).
Lemma 47. Let F : T → C be a homological functor. The following assertions

are equivalent:
(1) F is I-exact;
(2) L0 F (A) ∼
= F (A) and Lp F (A) = 0 for all p > 0, A ∈∈ T;
(3) L0 F (A) ∼
= F (A) for all A ∈∈ T.
The analogous assertions for contravariant functors are equivalent as well.
Proof. If F is I-exact, then F maps I-exact chain complexes in T to

exact chain complexes in C. This applies to I-projective resolutions, so
that (1)=⇒(2)=⇒(3). It follows from (3) and Lemma 46 that F maps
I-epimorphisms to epimorphisms. Since this characterises I-exact functors,
we get (3)=⇒(1).
It can happen that Lp F = 0 for all p > 0 although F is not I-exact.

We have a natural transformation L0 F (A) → F (A) (or F (A) → R0 F (A)),
which is induced by the augmentation map P• → A for an I-projective resolu-
tion. Lemma 47 shows that these maps are usually not bijective, although this
happens frequently for derived functors on Abelian categories.
Definition 48. We let ExtnT,I (A, B) be the nth right derived functor with respect
to I of the contravariant functor A
→ T(A, B).
We have natural maps T(A, B) → Ext0T,I (A, B), which usually are not
invertible. Lemma 46 yields long exact sequences
· · · ← ExtnT,I (A, D) ← ExtnT,I (B, D) ← ExtnT,I (C, D) ← Extn−1

T,I (A, D) ←
· · · ← Ext1T,I (C, D) ← Ext0T,I (A, D) ← Ext0T,I (B, D) ← Ext0T,I (C, D) ← 0
for any I-exact exact triangle A → B → C → A and any D ∈∈ T.

We claim that there are similar long exact sequences
0 → Ext0T,I (D, A) → Ext0T,I (D, B) → Ext0T,I (D, C) → Ext1T,I (D, A) → · · ·

T,I (D, C) → ExtT,I (D, A) → ExtT,I (D, B) → ExtT,I (D, C) → · · ·
→ Extn−1 n n n
in the second variable. Since P (D)n is I-projective, the sequences
0 → T(P (D)n , A) → T(P (D)n , B) → T(P (D)n , C) → 0
are exact for all n ∈ N. This extension of chain complexes yields the desired
long exact sequence.
We list a few more elementary properties of derived functors. We only spell
things out for the left derived functors Ln F : T → C of a covariant functor
F : T → C. Similar assertions hold for right derived functors of contravariant
functors.
The derived functors Ln F satisfy I ⊆ ker Ln F and hence descend to
functors Ln F : T/I → C because the zero map P (A) → P (B) is a chain
map lifting of f if f ∈ I(A, B). As a consequence, Ln F (A) ∼ = 0 if A
is I-contractible. The long exact homology sequences of Lemma 46
show that Ln F (f ) : Ln F (A) → Ln F (B) is invertible if f ∈ T(A, B) is an
I-equivalence.
Warning 49. The derived functors Ln F are not homological and therefore
do not deserve to be called I-exact even though they vanish on I-phantom
maps. Lemma 46 shows that these functors are only half-exact on I-exact
triangles. Thus Ln F (f ) need not be monic (or epic) if f is I-monic (or I-epic).
The problem is that the I-projective resolution functor P : T → Ho(T) is not
exact – it even fails to be stable.
The following remarks require a more advanced background in homological

algebra and are not going to be used in the sequel.
Remark 50. The derived functors introduced above, especially the Ext func-
tors, can be interpreted in terms of derived categories.
We have already observed in §3.2.1 that the I-exact chain complexes
form a thick subcategory of Ho(T). The augmentation map P (A) → A of
an I-projective resolution of A ∈∈ T is a quasi-isomorphism with respect to
this thick subcategory. The chain complex P (A) is projective (see [12]), that
is, for any chain complex C• , the space of morphisms A → C• in the derived
category Der(T, I) agrees with [P (A), C• ]. Especially, ExtnT,I (A, B) is the
space of morphisms A → B[n] in Der(T, I).
Now let F : T → C be an additive covariant functor. Extend it to an exact
functor F̄ : Ho(T) → Ho(C). It has a total left derived functor

LF̄ : Der(T, I) → Der(C), A
→ F̄ P (A) .

By definition, we have Ln F (A) := Hn LF̄ (A) .
Remark 51. In classical Abelian categories, the Ext groups form a graded ring,
and the derived functors form graded modules over this graded ring. The same
happens in our context. The most conceptual construction of these products
uses the description of derived functors sketched in Remark 50.
Recall that we may view elements of ExtnT,I (A, B) as morphisms A → B[n]
in the derived category Der(T, I). Taking translations, we can also view them
as morphisms A[m] → B[n + m] for any m ∈ Z. The usual composition in the
category Der(T, I) therefore yields an associative product
ExtnT,I (B, C) ⊗ Extm

T,I (A, B) → ExtT,I (A, C).
n+m
Thus
n we get a graded additive category with morphism spaces
ExtT,I (A, B) n∈N .
Similarly, if F : T → C is an additive functor and LF̄ : Der(T, I) → Der(C)
is as in Remark 51, then a morphism A → B[n] in Der(T, I) induces a mor-
phism LF̄ (A) → LF̄ (B)[n] in Der(C). Passing to homology, we get canonical
maps

ExtnT,I (A, B) → HomC LFm (A), LFm−n (B) ∀m ≥ n,
which satisfy an appropriate associativity condition. For a contravariant functor,

we get canonical maps

ExtnT,I (A, B) → HomC RF m (B), RF m+n (A) ∀m ≥ 0.
3.6. Projective objects via adjointness

We develop a method for constructing enough projective objects. Let T and C
be stable additive categories, let F : T → C be a stable additive functor, and
let I := ker F . In our applications, T is triangulated and the functor F is either
exact or stable and homological.
Recall that a covariant functor R : T → Ab is (co)representable if it is
naturally isomorphic to T(A, ) for some A ∈∈ T, which is then unique. If
the functor B
→ C A, F (B) on T is representable, we write F ( (A) for the
representing object. By construction, we have natural isomorphisms

T(F ( (A), B) ∼ = C A, F (B)
for all B ∈ T. Let C be the full subcategory of all objects A ∈∈ C for which
F ( (A) is defined. Then F ( is a functor C → T, which we call the (partially
defined) left adjoint of F . Although one usually assumes C = C , we shall also
need F ( in cases where it is not defined
everywhere.
The functor B
→ C A, F (B) for A ∈∈ C vanishes on I = ker F for triv-
ial reasons. Hence F ( (A) ∈∈ T is I-projective. This simple observation is
surprisingly powerful: as we shall see, it often yields all I-projective objects.
Remark 52. We have F ( (A) ∼= F ( (A) for all A ∈∈ C , so that (C ) = C .

(
Moreover, F commutes with infinite direct sums (as far as they exist in T)
because
. /
T F ( (Ai ), B ∼= T(F ( (Ai ), B) ∼= C Ai , F (B)
. /
∼
=C Ai , F (B) .
Example 53. Consider the functor K∗ : KK → AbZ/2 . Let Z ∈∈ AbZ/2 denote

the trivially graded Abelian group Z. Notice that

Hom Z, K∗ (A) ∼ = K0 (A) ∼
= KK(C, A),

Hom Z[1], K∗ (A) ∼ = K1 (A) ∼
= KK(C0 (R), A),
where Z[1] means Z in odd degree. Hence K(∗ (Z) = C and K(∗ (Z[1]) = C0 (R).
More generally, Remark 52 shows that K(∗ (A) is defined if both the even and
odd parts of A ∈∈ AbZ/2 are countable free Abelian groups: it is a direct sum
of at most countably many copies of C and C0 (R). Hence all such countable
direct sums are IK -projective (we briefly say K-projective). As we shall see,
K(∗ is not defined on all of AbZ/2 ; this is typical of homological functors.
Example 54. Consider the functor H∗ : Ho(C; Z/p) → CZ/p of Example 9.

Let j : CZ/p → Ho(C; Z/p) be the functor that views an object of CZ/p as a
p-periodic chain complex whose boundary map vanishes.
A chain map j (A) → B• for A ∈∈ CZ/p and B• ∈∈ Ho(C; Z/p) is a family
of maps ϕn : An → ker(dn : Bn → Bn−1 ). Such a family is chain homotopic to 0
if and only if each ϕn lifts to a map An → Bn+1 . Suppose that An is projective
for all n ∈ Z/p. Then such a lifting exists if and only if ϕn (An ) ⊆ dn+1 (Bn+1 ).
Hence

[j (A), B• ] ∼
= C An , Hn (B• ) ∼= CZ/p A, H∗ (B• ) .
n∈Z/p
As a result, the left adjoint of H∗ is defined on the subcategory of projective

objects P(C)Z/p ⊆ CZ/p and agrees there with the restriction of j . We will
show in §3.8 that all IH -projective objects are of the form H(∗ (A) for some
A ∈∈ P(C)Z/p (provided C has enough projective objects).
By duality, analogous results hold for injective objects: the domain of the
right adjoint of H∗ is the subcategory of injective objects of CZ/p , the right
adjoint is equal to j on this subcategory, and this provides all H∗ -injective
objects of Ho(C; Z/p).
These examples show that F ( yields many ker F -projective objects. We

want to get enough ker F -projective objects in this fashion, assuming that F (
'
is defined on enough of C. In order to treat ideals of the form Fi , we
now consider a more complicated setup. Let {Ci | i ∈ I } be a set of stable
homological or triangulated categories together with full subcategories PCi ⊆
Ci and stable homological or exact functors Fi : T → Ci for all i ∈ I . Assume
that
– the left adjoint Fi( is defined on PCi for all i ∈ I ;

– there is an epimorphism P → Fi (A) in Ci with P ∈∈ PCi for any i ∈ I ,
A ∈∈ T;

– the set of functors Fi( : PCi → T is cointegrable, that is, i∈I Fi( (Bi ) exists
for all families of objects Bi ∈ PCi , i ∈ I .
The reason for the notation PCi is that for a homological functor Fi we usually
take PCi to be the class of projective objects of Ci ; if Fi is exact, then we often
take PCi = Ci . But it may be useful to choose a smaller category, as long as it
satisfies the second condition above.
Proposition 55. In this situation, there are enough I-projective objects, and PI

is generated by i∈I {Fi( (B) | B ∈ PCi }. More precisely, an object of T is

I-projective if and only if it is a retract of i∈I Fi( (Bi ) for a family of objects
Bi ∈ PCi .

Proof. Let P̃0 := i∈I {Fi( (B) | B ∈ PCi } and P0 := (P̃0 )⊕ . To begin with,
we observe that any object of the form Fi( (B) with B ∈∈ PCi is
ker Fi -projective and hence I-projective because I ⊆ ker Fi . Hence P0 consists
of I-projective objects.
Let A ∈∈ T. For each i ∈ I , there is an epimorphism pi : Bi → Fi (A) with

Bi ∈ PCi . The direct sum B := i∈I Fi( (Bi ) exists. We have B ∈∈ P0 by
construction. We are going to construct an I-epimorphism p : B → A. This
shows that there are enough I-projective objects.
The maps pi : Bi → Fi (A) provide maps p̂i : Fi( (B i ) → A via the
adjointness isomorphisms T(Fi( (Bi ), A) ∼ = Ci Bi , Fi (A) . We let p :=
(
p̂i : Fi (Bi ) → A. We must check that p is an I-epimorphism. Equiva-
lently, p is ker Fi -epic for all i ∈ I ; this is, in turn equivalent to Fi (p) being an
epimorphism in Ci for all i ∈ I , because of Lemma 22 or 25. This is what we
are going to prove.
The identity map on Fi((Bi ) yields a map (
αi : Bi → F(i Fi (B i ) via the
adjointness isomorphism T Fi (Bi ), Fi (Bi ) ∼
( (
= Ci Bi , Fi Fi (Bi ) . Compos-
ing with the map
. /
Fi Fi( (Bi ) → Fi Fi( (Bi ) = Fi (B)
induced by the coordinate embedding Fi( (Bi ) → B, we get a map αi : Bi →

Fi (B). The naturality of the adjointness isomorphisms yields Fi (p̂i ) ◦ αi = pi
and hence Fi (p) ◦ αi = pi . The map pi is an epimorphism by assumption. Now
we use a cancellation result for epimorphisms: if f ◦ g is an epimorphism, then
so is f . Thus Fi (p) is an epimorphism as desired.
If A is I-projective, then the I-epimorphism p : B → A splits; to see this,
embed p in an exact triangle N → B → A → N and observe that the map
A → N belongs to I(A, N) = 0. Therefore, A is a retract of B. Since P0
is closed under retracts and B ∈∈ P0 , we get A ∈∈ P0 . Hence P̃0 generates
all I-projective objects.
3.7. The universal exact homological functor

For the following results, it is essential to define an ideal by a single functor F
instead of a family of functors as in Proposition 55.
Definition 56. Let I be a homological ideal in a triangulated category T. An
I-exact stable homological functor F : T → C is called universal if any other
I-exact stable homological functor G : T → C factors as Ḡ = G ◦ F for a
stable exact functor Ḡ : C → C that is unique up to natural isomorphism.
This universal property characterises F uniquely up to natural isomorphism.
We have constructed such a functor in §3.2.1. Beligiannis constructs it in
[4, §3] using a localisation of the Abelian category Coh(T) at a suitable Serre
subcategory; he calls this functor projectivisation functor and its target category
Steenrod category. This notation is motivated by the special case of the Adams
spectral sequence. The following theorem allows us to check whether a given
functor is universal:
Theorem 57. Let T be a triangulated category, let I ⊆ T be a homological
ideal, and let F : T → C be an I-exact stable homological functor into a stable
Abelian category C; let PC be the class of projective objects in C. Suppose that
idempotent morphisms in T split.
The functor F is the universal I-exact stable homological functor and there
are enough I-projective objects in T if and only if
– C has enough projective objects;
– the adjoint functor F ( is defined on PC;
– F ◦ F ( (A) ∼
= A for all A ∈∈ PC.
Proof. Suppose first that F is universal and that there are enough I-projective
objects. Then F is equivalent to the projectivisation functor of [4]. The various
properties of this functor listed in [4, Proposition 4.19] include the following:
– there are enough projective objects in C;
– F induces an equivalence of categories PI ∼ = PC (PI is the class of pro-
jective
objects in
T);
– C F (A), F (B) ∼ = T(A, B) for all A ∈∈ PI , B ∈∈ T.
Here we use the assumption that idempotents in T split. The last property is
equivalent to F ( ◦ F (A) ∼
= A for all A ∈∈ PI . Since PI ∼ = PC via F , this
( ( ∼
implies that F is defined on PC and that F ◦ F (A) = A for all A ∈∈ PC.
Thus F has the properties listed in the statement of the theorem.
Now suppose conversely that F has these properties. Let PI ⊆ T be
the essential range of F ( : PC → T. We claim that PI is the class of all
I-projective objects in T. Since F ◦ F ( is equivalent to the identity functor

on PC by assumption, F |PI and F ( provide an equivalence of categories
PI ∼
= PC. Since C is assumed to have enough projectives, the hypotheses
of Proposition 55 are satisfied. Hence there are enough I-projective objects
in T, and any object of PI is a retract of an object of PI . Idempotent mor-
phisms in the category PI ∼ = PC split because C is Abelian and retracts of
projective objects are again projective. Hence PI is closed under retracts
in T, so that PI = PI . It also follows that F and F ( provide an equiva-
lence of categories
PI ∼= PC. Hence F ( ◦ F (A) ∼ = A for all A ∈∈ PI , so
that we get C F (A), F (B) ∼ = T(F (
◦ F (A), B) ∼
= T(A, B) for all A ∈∈ PI ,
B ∈∈ T.
Now let G : T → C be a stable homological functor. We will later assume G
to be I-exact, but the first part of the following argument works in general.

Since F provides an equivalence of categories PI ∼ = PC, the rule Ḡ F (P ) :=
G(P ) defines a functor Ḡ on PC. This yields a functor Ho(Ḡ) : Ho(PC) →
Ho(C ). Since C has enough projective objects, the construction of projective
resolutions provides a functor P : C → Ho(PC). We let Ḡ be the composite
functor
P Ho(Ḡ) H0
→ Ho(PC) −−−→ Ho(C ) −→ C .
Ḡ : C −
This functor is right-exact and satisfies Ḡ ◦ F = G on I-projective objects

of T.
Now suppose that G is I-exact. Then we get Ḡ ◦ F = G for all objects
of T because this holds for I-projective objects. We claim that Ḡ is exact.
Let A ∈∈ C. Since C has enough projective objects, we can find a projective
resolution of A. We may assume this resolution to have the form F (P• ) with
P• ∈∈ Ho(PI ) because F (PI ) ∼ = PC. Lemma 28 yields that P• is I-exact
except in degree 0. Since I ⊆ ker G, the chain complex P• is ker G-exact in
positive degrees as well, so that G(P• ) is exact except in degree 0 by Lemma 28.
As a consequence, Lp Ḡ(A) = 0 for all p > 0. We also have L0 Ḡ(A) = Ḡ(A)
by construction. Thus Ḡ is exact.
As a result, G factors as G = Ḡ ◦ F for an exact functor Ḡ : C → C . It is
clear that Ḡ is stable. Finally, since C has enough projective objects, a functor
on C is determined up to natural equivalence by its restriction to projective
objects. Therefore, our factorisation of G is unique up to natural equivalence.
Thus F is the universal I-exact functor.
Remark 58. Let P C ⊆ PC be some subcategory such that any object of C

is a quotient of a direct sum of objects of P C. Equivalently, (P C)⊕ = PC.
Theorem 57 remains valid if we only assume that F ( is defined on P C and
that F ◦ F ( (A) ∼
= A holds for A ∈∈ P C because both conditions are evidently
hereditary for direct sums and retracts.
Theorem 59. In the situation of Theorem 57, the domain of the functor F (
contains PC, and its essential range is PI . The functors F and F ( restrict to
equivalences of categories PI ∼= PC inverse to each other.
An object A ∈∈ T is I-projective if and only if F (A) is projective and

C F (A), F (B) ∼ = T(A, B)
for all B ∈∈ T; following Ross Street [24], we call such objects F -projective.
We have F (A) ∈∈ PC if and only if there is an I-equivalence P → A with
P ∈∈ PI .
The functors F and F ( induce bijections between isomorphism classes of
projective resolutions of F (A) in C and isomorphism classes of I-projective
resolutions of A ∈∈ T in T.
If G : T → C is any (stable) additive functor, then there is a unique
right-exact (stable) functor Ḡ : C → C such that Ḡ ◦ F (P ) = G(P ) for all
P ∈∈ PI .
The left derived functors of G with respect to I and of Ḡ are related by
natural isomorphisms Ln Ḡ ◦ F (A) = Ln G(A) for all A ∈∈ T, n ∈ N. There
is a similar statement for cohomological functors, which specialises to natural
isomorphisms

ExtnT,I (A, B) ∼
= ExtnC F (A), F (B) .
Proof. We have already seen during the proof of Theorem 57 that F

∼
=
restricts
to an equivalence
of categories P I −
→ PC with inverse F ( and that
C F (A), F (B) ∼ = T(A, B) for all A ∈∈ PI , B ∈∈ PC.
Conversely, if A is F -projective
in the sense
of Street, then A is I-projective
because already T(A, B) ∼ = C F (A), F (B) for all B ∈∈ T yields A ∼ = F( ◦
F (A), so that A is I-projective; notice that the projectivity of F (A) is automatic.
Since F maps I-equivalences to isomorphisms, F (A) is projective whenever
there is an I-equivalence P → A with I-projective P . Conversely, suppose
that F (A) is I-projective. Let P0 → A be a one-step I-projective resolution.
Since F (A) is projective, the epimorphism F (P0 ) → F (A) splits by some map
F (A) → F (P0 ). The resulting map F (P0 ) → F (A) → F (P0 ) is idempotent
and comes from an idempotent endomorphism of P0 because F is fully faithful
on PI . Its range object P exists because we require idempotent morphisms
in C to split. It belongs again to PI , and the induced map F (P ) → F (A) is
invertible by construction. Hence we get an I-equivalence P → A.
If C• is a chain complex over T, then we know already from Lemma 28

that C• is I-exact if and only if F (C• ) is exact. Hence F maps an I-projective
resolution of A to a projective resolution of F (A). Conversely, if P• → F (A)
is any projective resolution in C, then it is of the form F (P̂• ) → F (A) where
P̂• := F ( (P• ) and where we get the map P̂0 → A by adjointness from the given
map P0 → F (A). This shows that F induces a bijection between isomorphism
classes of I-projective resolutions of A and projective resolutions of F (A).
We have seen during the proof of Theorem 57 how a stable homological
functor G : T → C gives rise to a unique right-exact functor Ḡ : C →C that
satisfies Ḡ ◦ F (P ) = G(P ) for all P ∈∈ PI . The derived functors Ln Ḡ F (A)
for A ∈∈ T are computed by applying Ḡ to a projective resolution of F (A).
Since such a projective resolution is of the form F (P• ) for an I-projective
resolution P• → A and since Ḡ ◦F = G on I-projective objects, the derived
functors Ln G(A) and Ln Ḡ F (A) are computed by the same chain complex
and agree. The same reasoning applies to cohomological functors and yields
the assertion about Ext.
If F ( (X) is defined, then it can be shown that X ∼ = X0 ⊕ X1 where

X1 ∼
= F F ( (X) is projective and F ( (X0 ) = 0, that is, C(X0 , F Y ) = 0 for all Y .
Usually, the latter condition implies X0 = 0, but it is not clear whether this
implication is true in general.
Remark 60. The assumption that idempotents split is only needed to check
that the universal I-exact functor has the properties listed in Theorem 57.
The converse directions of Theorem 57 and Theorem 59 do not need this
assumption.
If T has countable direct sums or countable direct products, then idempotents
in T automatically split by [19, §1.3]. This covers categories such as KKG
because they have countable direct sums.
3.8. Derived functors in homological algebra

Now we study the kernel IH of the homology functor H∗ : Ho(C; Z/p) →
CZ/p introduced in Example 9. We get exactly the same statements if we
replace the homotopy category by its derived category and study the ker-
nel of H∗ : Der(C; Z/p) → CZ/p . We often abbreviate IH to H and speak of
H-epimorphisms, H-exact chain complexes, H-projective resolutions, and so
on. We denote the full subcategory of H-projective objects in Ho(C; Z/p)
by PH .
We assume that the underlying Abelian category C has enough projective
objects. Then the same holds for CZ/p , and we have P(CZ/p ) ∼= (PC)Z/p .
That is, an object of CZ/p is projective if and only if its homogeneous pieces
are.
Theorem 61. The category Ho(C; Z/p) has enough H-projective objects, and
the functor H∗ : Ho(C; Z/p) → CZ/p is the universal H-exact stable homo-
logical functor. Its restriction to PH provides an equivalence of categories
PH ∼ = PCZ/p . More concretely, a chain complex in Ho(C; Z/p) is H-projective
if and only if it is homotopy equivalent to one with vanishing boundary map
and projective entries.
The functor H∗ maps isomorphism classes of H-projective resolutions of a
chain complex A in Ho(C; Z/p) bijectively to isomorphism classes of projective
resolutions of H∗ (A) in CZ/p . We have

ExtnHo(C;Z/p),IH (A, B) ∼
= ExtnC H∗ (A), H∗ (B) .
Let F : C → C be some covariant additive functor and define
F̄ : Ho(C; Z/p) → Ho(C ; Z/p)

by applying F entrywise. Then Ln F̄ (A) ∼
= Ln F H∗ (A) for all n ∈ N. Simi-

larly, we have Rn F̄ (A) ∼
= Rn F H∗ (A) if F is a contravariant functor.
Proof. The category CZ/p has enough projective objects by assumption. We
have already seen in Example 54 that H(∗ is defined on PCZ/p ; this functor is
denoted by j in Example 54. It is clear that H∗ ◦ j (A) ∼
= A for all A ∈∈ CZ/p .
Now Theorem 57 shows that H∗ is universal. We do not need idempotent
morphisms in Ho(C; Z/p) to split by Remark 60.
Remark 62. Since the universal I-exact functor is essentially unique, the uni-
versality of H∗ : Der(C; Z/p) → CZ/p means that we can recover this functor
and hence the stable Abelian category CZ/p from the ideal IH ⊆ Der(C; Z/p).
That is, the ideal IH and the functor H∗ : Der(C; Z/p) → CZ/p contain exactly
the same amount of information.
For instance, if we forget the precise category C by composing H∗ with some
faithful functor C → C , then the resulting homology functor Ho(C; Z/p) → C
still has kernel IH . We can recover CZ/p by passing to the universal I-exact
functor.
We compare this with the situation for truncation structures ([3]). These
cannot exist for periodic categories such as Der(C; Z/p) for p ≥ 1. Given the
standard truncation structure on Der(C), we can recover the Abelian category C
as its core; we also get back the homology functors Hn : Der(C) → C for all
n ∈ Z. Conversely, the functor H∗ : Der(C) → CZ together with the grading
on CZ tells us what it means for a chain complex to be exact in degrees
greater than 0 or less than 0 and thus determines the truncation structure.
Hence the standard truncation structure on Der(C) contains the same amount
of information as the functor H∗ : Der(C) → CZ together with the grading
on CZ .
4. The plain Universal Coefficient Theorem

Now we study the ideal IK := ker K∗ ⊆ KK of Example 5. We complete our
analysis of this example and explain the Universal Coefficient Theorem for KK
in our framework. We call IK -projective objects and IK -exact functors briefly
K-projective and K-exact and let PK be the class of K-projective objects in
KK.
Let AbZ/2
c ⊆ AbZ/2 be the full subcategory of countable Z/2-graded Abelian
groups. Since the K-theory of a separable C∗ -algebra is countable, we may
view K∗ as a stable homological functor K∗ : KK → AbZ/2 c .
Theorem 63. There are enough K-projective objects in KK, and the universal
K-exact functor is K∗ : KK → AbZ/2
c . It restricts to an equivalence of categories
Z/2
between PK and the full subcategory Abfc ⊆ AbZ/2 c of Z/2-graded countable
∗
free Abelian groups. A separable C -algebra belongs to PK if and only if it

is KK-equivalent to i∈I0 C ⊕ i∈I1 C0 (R) where the sets I0 , I1 are at most
countable.
If A ∈∈ KK, then K∗ maps isomorphism classes of K-projective resolutions
of A in T bijectively to isomorphism classes of free resolutions of K∗ (A). We
have
⎧
⎪
⎪ K∗ (A), K∗ (B) for n = 0;
⎨HomAbZ/2
ExtnKK,IK (A, B) =∼ Ext1 Z/2 K∗ (A), K∗ (B) for n = 1;
⎪
⎪ Ab
⎩0 for n ≥ 2.
Let F : KK → C be some covariant additive functor; then there is a

unique right-exact functor F̄ : AbZ/2
c → C with F̄ ◦ K∗ = F . We have Ln F =
(Ln F̄ ) ◦ K∗ for all n ∈ N; this vanishes for n ≥ 2. Similar assertions hold for
contravariant functors.
Proof. Notice that AbZ/2 c ⊆ AbZ/2 is an Abelian category. We shall denote

objects of AbZ/2 by pairs (A0 , A1 ) of Abelian groups. By definition,
Z/2
(A0 , A1 ) ∈∈ Abfc if and only if A0 and A1 are countable free Abelian groups,
that is, they are of the form A0 = Z[I0 ] and A1 = Z[I1 ] for at most countable
sets I0 , I1 . It is well-known that any Abelian group is a quotient of a free Abelian
group and that subgroups of free Abelian groups are again free. Moreover, free
Z/2
Abelian groups are projective. Hence Abfc is the subcategory of projective
objects in AbZ/2
c and any object G ∈∈ AbZ/2
c has a projective resolution of
Z/2
the form 0 → F1 → F0 G with F0 , F1 ∈∈ Abfc . This implies that derived
Z/2
functors on Abc only occur in dimensions 1 and 0.
Z/2
As in Example 53, we see that K(∗ is defined on Abfc and satisfies

K(∗ Z[I0 ], Z[I1 ] ∼
= C⊕ C0 (R)
i∈I0 i∈I1

if I0 , I1 are countable. We also have K∗ ◦ K(∗ Z[I0 ], Z[I1 ] ∼
= Z[I0 ], Z[I1 ] ,
so that the hypotheses of Theorem 57 are satisfied. Hence there are enough
K-projective objects and K∗ is universal. The remaining assertions follow
from Theorem 59 and our detailed knowledge of the homological algebra in
AbZ/2
c .
Example 64. Consider the stable homological functor
F : KK → AbZ/2
c , A
→ K∗ (A ⊗ B)
for some B ∈∈ KK, where ⊗ denotes, say, the spatial C∗ -tensor product. We
claim that the associated right-exact functor AbZ/2
c → AbZ/2
c is
F̄ : AbZ/2
c → AbZ/2
c , G
→ G ⊗ K∗ (B).
It is easy to check F ◦ K(∗ (G) ∼

= G ⊗ K∗ (B) ∼
Z/2
= F̄ (G) for G ∈∈ Abfc . Since
the functor G
→ G ⊗ K∗ (B) is right-exact and agrees with F̄ on projective
objects, we get F̄ (G) ∼= G ⊗ K∗ (B) for all G ∈∈ AbZ/2c . Hence the derived
functors of F are
⎧
⎪
⎪K∗ (A) ⊗ K∗ (B) for n = 0;
⎨
∼
Ln F (A) = Tor K∗ (A), K∗ (B) for n = 1;
1
⎪
⎪
⎩0 for n ≥ 2.
Here we use the same graded version of Tor as in the Künneth Theorem ([5]).
Example 65. Consider the stable homological functor
F : KK → AbZ/2 , B
→ KK∗ (A, B)
for some A ∈∈ KK. We suppose that A is a compact

object of KK, that is, the
functor F commutes with direct sums. Then KK∗ A, K∗ (G) ∼
(
= KK∗ (A, C) ⊗
Z/2
G for all G ∈∈ Abfc because this holds for G = (Z, 0) and is inherited by
suspensions and direct sums. Now we get F̄ (G) ∼
= KK∗ (A, C) ⊗ G for all
G ∈∈ AbZ/2
c as in Example 64. Therefore,
⎧
⎪
⎨KK∗(A, C) ⊗ K∗ (B) for n = 0;
⎪
∼
Ln F (B) = Tor1 KK∗ (A, C), K∗ (B) for n = 1;
⎪
⎪
⎩0 for n ≥ 2.
Generalising Examples 64 and 65, we have F̄ (G) ∼

= F (C) ⊗ G and hence
⎧
⎪
⎨F (C) ⊗ K∗ (B) for n = 0,
⎪
∼
Ln F (B) = Tor1 F (C), K∗ (B) for n = 1,
⎪
⎪
⎩0 for n ≥ 2,
for any covariant functor F : KK → C that commutes with direct sums.

Similarly, if F : KK → C is contravariant and maps direct sums to direct
products, then F̄ (G) ∼= Hom(G, F (C)) and
⎧
⎪
⎨Hom K∗ (B), F (C) for n = 0,
⎪
Rn F (B) ∼
= Ext1 K∗ (B), F (C) for n = 1,
⎪
⎪
⎩0 for n ≥ 2.
The description of ExtnKK,IK in Theorem 63 is a special case of this.
4.1. Universal Coefficient Theorem in the hereditary case

In general, we need spectral sequences in order to relate the derived func-
tors Ln F back to F . We will discuss this in a sequel to this article. Here we
only treat the simple case where we have projective resolutions of length 1.
The following universal coefficient theorem is very similar to but slightly more
general than [4, Theorem 4.27] because we do not require I-equivalences to be
invertible.
Theorem 66. Let I be a homological ideal in a triangulated category T.

Let A ∈∈ T have an I-projective resolution of length 1. Suppose also that
T(A, B) = 0 for all I-contractible B. Let F : T → C be a homological functor,
F̃ : Top → C a cohomological functor, and B ∈∈ T. Then there are natural
short exact sequences
0 → L0 F∗ (A) → F∗ (A) → L1 F∗−1 (A) → 0,

0 → R1 F̃ ∗−1 (A) → F̃ ∗ (A) → R0 F̃ ∗ (A) → 0,
0 → Ext1T,I (A, B) → T(A, B) → Ext0T,I (A, B) → 0.
Example 67. For the ideal IK ⊆ KK, any object has a K-projective resolution
of length 1 by Theorem 63. The other hypothesis of Theorem 66 holds if and
only if A satisfies the Universal Coefficient Theorem (UCT). The UCT for
KK(A, B) predicts KK(A, B) = 0 if K∗ (B) = 0. Conversely, if this is the case,
then Theorem 66 applies, and our description of ExtKK,IK in Theorem 63 yields
the UCT for KK(A, B) for all B. This yields our claim.
Thus the UCT for KK(A, B) is a special of Theorem 66. In the situa-
tions of Examples 64 and 65, we get the familiar Künneth Theorems for
K∗ (A ⊗ B) and KK∗ (A, B). These arguments are very similar to the original
proofs (see [5]). Our machinery allows us to treat other situations in a similar
fashion.
Proof of Theorem 66. We only write down the proof for homological func-
tors. The cohomological case is dual and contains T( , B) as a special
case.
δ1 δ0
Let 0 → P1 −→ P0 −→ A be an I-projective resolution of length 1 and view
it as an I-exact chain complex of length 3. Lemma 31 yields a commuting
diagram
P1
δ1
/ P0 δ̃0
/ Ã
α

P1
δ1
/ P0 δ0
/ A,
such that the top row is part of an I-exact exact triangle P1 → P0 → Ã → P1
and α is an I-equivalence. We claim that α is an isomorphism in T.
α β
We embed α in an exact triangle −1 B → Ã −
→A− → B. Lemma 24 shows
that B is I-contractible because α is an I-equivalence. Hence T(A, B) = 0
by our assumption on A. This forces β = 0, so that our exact triangle splits:
Ã ∼
= A ⊕ −1 B. Now we apply the functor T(·, B) to the exact triangle P0 →
P1 → Ã. The resulting long exact sequence has the form
· · · ← T(P0 , B) ← T(Ã, B) ← T(P1 , B) ← · · · .
Since both P0 and P1 are I-projective and B is I-contractible, we get T(Ã, B) =

0. Then T(B, B) ⊆ T(Ã, B) vanishes as well, so that B ∼ = 0 and α is invertible.
δ1 δ0
We get an exact triangle in T of the form P1 − → P0 −→ A → P1 because
any triangle isomorphic to an exact one is itself exact.
Now we apply F . Since F is homological, we get a long exact sequence
F∗ (δ1 ) F∗−1 (δ1 )
· · · → F∗ (P1 ) −−−→ F∗ (P0 ) → F∗ (A) → F∗−1 (P1 ) −−−−→ F∗−1 (P0 ) → · · · .
We cut this into short exact sequences of the form

coker F∗ (δ1 ) F∗ (A) ker F∗−1 (δ1 ) .
Since coker F∗ (δ1 ) = L0 F∗ (A) and ker F∗ (δ1 ) = L1 F∗ (A), we get the desired
exact sequence. The map L0 F∗ (A) → F∗ (A) is the canonical map induced
by δ0 . The other map F∗ (A) → L1 F∗−1 (A) is natural for all morphisms between
objects with an I-projective resolution of length 1 by Proposition 44.
The proof shows that – in the situation of Theorem 66 – we have
Ext0T,I (A, B) ∼
= T/I(A, B), Ext1T,I (A, B) ∼
= I(A, B).
More generally, we can construct a natural map I(A, B) → Ext1T,I (A, B) for
any homological ideal, using the I-universal homological functor F : T → C.
We embed f ∈ I(A, B) in an exact triangle B → C → A → B. We get
an extension
* +
F (B) F (C) F (A) ∈∈ Ext1C F (A), F (B)

because this triangle is I-exact. This class κ(f ) in Ext1C F (A), F (B) does
not depend on auxiliary choices because the exact triangle B → C →
A → B is unique ∼
up to isomorphism. Theorem 59 yields ExtT,I (A, B) =
1
1
ExtC F (A), F (B) because F is universal. Hence we get a natural map
κ : I(A, B) → Ext1T,I (A, B).
We may view κ as a secondary invariant generated by the canonical map
T(A, B) → Ext0T,I (A, B).
For the ideal IK , we get the same map κ as in Example 5.

An Abelian category with enough projective objects is called hereditary
if any subobject of a projective object is again projective. Equivalently, any
object has a projective resolution of length 1. This motivates the following
definition:
Definition 68. A homological ideal I in a triangulated category T is called

hereditary if any object of T has a projective resolution of length 1.
If I is hereditary and if I-equivalences are invertible, then Theorem 66

applies to all A ∈∈ T (and vice versa).
5. Crossed products for compact quantum groups

Let G be a compact (quantum) group and write C(G) and C∗ (G) for the Hopf
C∗ -algebras of Ĝ and G. We study the homological algebra in KKG generated
by the ideals I,K ⊆ I ⊆ KK defined in Example 7. Recall that I is the
kernel of the crossed product functor
G : KKG → KK, A
→ G A,
whereas I,K is the kernel of the composite functor K∗ ◦ (G ). Since we

have already analysed K∗ in §4, we can treat both ideals in a parallel fashion.
Our setup contains two classical special cases. First, G may be a compact
Lie group. Then C(G) and C∗ (G) have the usual meaning, and the objects
of KKG are separable C∗ -algebras with a continuous action of G. Secondly,
C(G) may be the dual C∗red (H ) of a discrete group H , so that C∗ (G) = C0 (H ).
Then the objects of KKG are separable C∗ -algebras equipped with a (reduced)
coaction of H . (We disregard the nuances between reduced and full coactions.)
If we identify KKCred (H ) ∼
∗
= KKH using Baaj–Skandalis duality, then the crossed
product functor G corresponds to the forgetful functor KKH → KK.
Let P and P,K be the classes of projective objects for the ideals I
and I,K . Our first task is to find enough projective objects for these two
ideals.
Let τ : KK → KKG be the functor that equips B ∈∈ KK with the triv-
ial G-action (that is, coaction of C(G)). This corresponds to the induc-
tion functor from the trivial quantum group to C∗ (G) under the equivalence
KKG ∼
∗
= KKC (G) .
The functors τ and G are adjoint, that is, there are natural isomorphisms
KKG (τ (A), B) ∼
= KK(A, G B) (5.1)
for all A ∈∈ KK, B ∈∈ KKG . This generalisation of the Green–Julg Theorem

is proved in [26, Théorème 5.10].
Lemma 69. There are enough projective objects for I and I,K . We have
P = (τ (KK))⊕ , P,K = (τ C, τ C0 (R))⊕ .
Proof. The adjoint of G is defined on all of KK, which is certainly enough

for Proposition 55. This yields the assertions for I and P ; even more, P
is the closure of τ (KK) under retracts.
The adjoint of K∗ ◦ (G ) is τ ◦ K(∗ , which is defined for free countable

Z/2-graded Abelian groups. Explicitly,

τ ◦ K(∗ Z[I0 ], Z[I1 ] ∼
= τ (C) ⊕ τ C0 (R) .
i∈I0 i∈I1
Z/2
Since any object of AbZ/2
c is a quotient of one in Abfc , Proposition 55 applies
and yields the assertions about I,K and P,K ; even more, P,K is the closure
Z/2
of τ ◦ K(∗ (Abfc ) under retracts.
To proceed further, we describe the universal I-exact functors. The functors

(G ) and K∗ ◦ (G ) fail the criterion of Theorem 57 (unless G = {1})
because
= C∗ (G) ⊗ A ∼
G τ (A) ∼ A.
=
This is not surprising because G is equivalent to a forgetful functor. The

universal functor recovers a linearisation of the forgotten structure.
First we consider the ideal I,K . By Lemma 69, the objects τ (C) and τ (C)
generate all I,K -projective objects. Their internal symmetries are encoded
by the Z/2-graded ring KKG ∗ (τ C, τ C) ; the superscript op denotes that we
op
take the product in reversed order. For classical compact groups, this ring is
commutative, so that the order of multiplication does not matter; in general,
the reversed-order product is the more standard choice. The following fact is
well-known:
Lemma 70. The ring KKG ∗ (τ C, τ C) is isomorphic to the representation ring

op
Rep(G) of G for ∗ = 0 and 0 for ∗ = 1.
We may take this as the definition of Rep(G).
Proof. The adjointness isomorphism (5.1) identifies

∼ ∗ ∼ ∗
KKG ∗ (τ C, τ C) = KK∗ (C, G τ C) = KK∗ C, C (G) = K∗ (C G).
Since G is compact, C∗ G is a direct sum of matrix algebras, one for each

irreducible representation of G. Hence the underlying Abelian group of Rep(G)
is the free Abelian group Z[Ĝ] on the set Ĝ of irreducible representations
of G. The ring structure on Rep(G) comes from the internal tensor product of
representations: represent two elements α, β of KKG 0 (τ C, τ C) by differences
of finite-dimensional representations π, $ of G, then α ◦ β ∈ KKG 0 (τ C, τ C) is
represented by $ ⊗ π because the product in KK boils down to an exterior
G
tensor product in this case (with reversed order).

Example 71. If C(G) = C∗red (H ) for a discrete group H , then Ĝ = H and the
product on Rep(G) = Z[H ] is the usual convolution. Thus Rep(G) is the group
ring of H .
If G is a compact group, then Rep(G) is the representation ring of G in the
usual sense.
For any B ∈∈ KKG , the Kasparov product turns KKG ∗ (τ C, B) into a left
op ∼
module over the ring KKG0 (τ C, τ C) = Rep(G). Thus KK ∗ (τ C, B) becomes
G
Z/2
an object of the Abelian category Mod(Rep G)c of Z/2-graded countable
Rep(G)-modules. We get a stable homological functor
FK : KKG → Mod(Rep G)Z/2

c , ∗ (τ C, B).
FK (B) := KKG
∗ (τ C, B) is
By (5.1), the underlying Abelian group of KKG
∼ ∼
∗ (τ C, B) = KK∗ (C, G B) = K∗ (G B).
KKG
Hence we still have ker FK = I,K . We often write FK (B) = K∗ (G B) if it is

Z/2
clear from the context that we view K∗ (G B) as an object of Mod(Rep G)c .
Theorem 72. The functor FK is the universal I,K -exact functor.

The subcategory of I,K -projective objects in KKG is equivalent to the sub-
category of Z/2-graded countable projective Rep(G)-modules. If A ∈∈ KKG ,
then FK induces a bijection between isomorphism classes of I,K -projective
Z/2
resolutions of A and projective resolutions of FK (A) in Mod(Rep G)c .
If A, B ∈∈ KKG , then

ExtnKKG ,I,K (A, B) ∼
= ExtnRep(G) K∗ (G A), K∗ (G B) .
If H : KKG → C is homological and commutes with countable direct sums,

then

Ln H (A) ∼
= TorRep(G)
n H∗ (τ C), K∗ (G A) ;
if H : (KKG )op → C is cohomological and turns countable direct sums into

direct products, then

Rn H (A) ∼
= ExtnRep(G) K∗ (G A), H ∗ (τ C) ;
Here we use the right or left Rep(G)-module structure on H∗ (τ C) that comes

from the functoriality of H .
Z/2
Proof. We verify universality using Theorem 57. The category Mod(Rep G)c
has enough projective objects: countable free modules are projective, and any
object is a quotient of a countable free module.
Given a free module (Rep(G)[I0 ], Rep(G)[I1 ]), we have natural isomor-

phisms

HomRep(G) (Rep(G)[I0 ], Rep(G)[I1 ]), FK (B) ∼= Kε (G B)
ε∈{0,1},i∈Iε

∼
= KKG τ C, B .
ε
ε∈{0,1},i∈Iε
Hence the adjoint functor FK( is defined on countable free modules. Idempotents
in KKG split by Remark 60. Therefore, the domain of FK( is closed under
Z/2
retracts and contains all projective objects of Mod(Rep G)c . It is easy to
see that FK ◦ FK( (A) ∼= A for free modules. This extends to retracts and hence
holds for all projective modules (compare Remark 58). Now Theorem 57 yields
that FK is universal.
The assertions about projective objects, projective resolutions, and Ext now
follow from Theorem 59. Theorem 59 also yields a formula for left derived
Z/2
functors in terms of the right-exact functor H̄ : Mod(Rep G)c → C associ-
ated to a homological functor H : KKG → C. It remains to compute H̄ .
First we define the Tor objects in the statement of the theorem if C is the
category of Abelian groups. Then H∗ (τ C) ∈∈ Mod(Rep G)Z/2 , and we can
take the derived functors of the usual Z/2-graded balanced tensor product
⊗Rep G for Rep(G)-modules. We claim that there are natural isomorphisms
H̄∗ (M) ∼
= H∗ (τ C) ⊗Rep G M
Z/2
for all M ∈∈ Mod(Rep G)c . This holds for M = Rep G and hence for all
free modules because we have natural isomorphisms

H̄∗ (Rep G) ∼
= H∗ FK( (Rep G) ∼= H∗ (τ C) ∼
= H∗ (τ C) ⊗Rep G Rep G.
For general M, the functor H̄ (M) is computed by a free resolution because it

is right-exact. Using this, one extends the computation to all modules. By defi-
Z/2
nition, TornRep G (N, M) for N ∈∈ Mod((Rep G)op )Z/2 , M ∈∈ Mod(Rep G)c ,
Z/2
is the nth left derived functor of the functor N ⊗Rep G on Mod(Rep G)c .
Now Theorem 59 yields the formula for Ln H provided H takes values in
Abelian groups. The same argument works in general, we only need more
complicated categories.
Let CZ/2 [Rep(G)op ] be the category of Z/2-graded objects A of C together
with a ring homomorphism Rep(G)op → CZ/2 (A, A). We can extend the defi-
nition of ⊗Rep G to get an additive stable bifunctor
⊗Rep G : CZ/2 [Rep(G)op ] ⊗ Mod(Rep G)Z/2 → CZ/2 .

Its derived functors are TornRep G . As above, we see that this yields the derived
functors of H .
The case of cohomological functors is similar and left to the reader.
If G is a compact group, then the same derived functors appear in the

Universal Coefficient Theorem for KKG by Jonathan Rosenberg and Claude
Schochet ([22]). This is no coincidence, of course. It is explained in [16] how a
homological ideal in a triangulated category generates a spectral sequence.
This machinery applied to the ideal I,K yields the spectral sequence of
[22].
In order to get the universal functor for the ideal I , we must lift the Rep G-
module structure on K∗ (G B) to G B. Given any additive category C,
we define a category C[Rep G] as in the proof of Theorem 72: its objects are
pairs (A, μ) where A ∈∈ C and μ is a ring homomorphism Rep(G) → C(A, A);
its morphisms are morphisms in C that are compatible with the Rep(G)-module
structure in the obvious sense.
A module structure μ : Rep(G) → KK(A, A) for A ∈∈ KK is equivalent
to a natural family of Rep(G)-module structures in the usual sense on the
groups KK(D, A) for all D ∈∈ KK: simply define x · y := μ(x) ◦ y for x ∈
Rep(G), y ∈ KK(D, A) and notice that this recovers the homomorphism μ for
y = idA .
The crucial property of the universal I,K -exact functor FK is that it lifts
the original functor K∗ (G ) : KKG → AbZ/2 c to a functor
FK : KKG → Mod(Rep G)Z/2

c = AbZ/2
c [Rep G].
We need a similar lifting of G : KKG → KK to KK[Rep G]. This requires

a simple special case of exterior products in KKG . In general, it is not so
easy to define exterior products in KKG for quantum groups because diagonal
actions on C∗ -algebras are not defined without additional structure. The only
case where this is easy is if one of the factors carries the trivial coaction. This
exterior product operation on the level of C∗ -algebras also works for Kasparov
cycles, that is, we get canonical maps
0 (A, B) ⊗ KK0 (C, D) → KK0 (A ⊗ C, B ⊗ D)

KKG G
for all A, B ∈∈ KKG , C, D ∈∈ KK. Equivalently, (A, C)

→ A ⊗ C is a
bifunctor KKG ⊗ KK → KKG .
This exterior product construction yields a natural map
= KKG (τ C, τ C) → KKG (τ C ⊗ A, τ C ⊗ A) ∼
$A : Rep(G)op ∼ = KKG (τ A, τ A)
for A ∈∈ KK, whose range commutes with the range of the map
τ : KK(A, A) → KKG (τ A, τ A).
The ring KKG (τ A, τ A) acts on KKG (τ A, B) ∼
= KK(A, G B) on the right
by Kasparov product. Hence so does Rep(G)op via $A . Thus KK(A, G B)
becomes a left Rep(G)-module for all A ∈∈ KK, B ∈∈ KKG . These module
structures are natural in the variable A because the images of KKG (τ C, τ C)
and KK(A, A) in KKG (τ A, τ A) commute. Hence they must come from a ring
homomorphism
μB : Rep(G) → KK(G B, G B).
These ring homomorphisms are natural because the Rep(G)-module structures
on KK(A, G B) are manifestly natural in B. Thus we have lifted G to a
functor
F : KKG → KK[Rep G], B
→ (G B, μB ).
It is clear that ker F = I . The target category KK[Rep G] is neither tri-
angulated nor Abelian. To remedy this, we use the Yoneda embedding
Y : KK → Coh(KK) constructed in §2.4. This embedding is fully faithful;
so is the resulting functor KK[Rep G] → Coh(KK)[Rep G].
Theorem 73. The functor Y ◦ F : KKG → Coh(KK)[Rep G] is the universal
I -exact functor.
Proof. We omit the proof of this theorem because it is only notationally more
difficult than the proof of Theorem 72.
The category Coh(KK)[Rep G] is not as terrible as it seems. We can usually
stay within the more tractable subcategory KK[Rep G], and many standard
techniques of homological algebra like bar resolutions work in this setting.
This often allows us to compute derived functors on Coh(KK)[Rep G] in more
classical terms.
Recall that the bar resolution of a Rep G-module M is a natural free resolu-
tion
· · · → (Rep G)⊗n ⊗ M → (Rep G)⊗n−1 ⊗ M → · · · → Rep G ⊗ M → M
with certain natural boundary maps. Defining

(Rep G)⊗n ⊗ M = Z[Ĝn ] ⊗ M := M,
x∈Ĝn
we can make sense of this in C[Rep G] provided C has countable direct sums;
the Rep G-module structures and the boundary maps can also be defined.
If A ∈∈ KK[Rep G], then the bar resolution lies in KK[Rep G] and is a

projective resolution of A in the ambient Abelian category Coh(KK)[Rep G].
Hence we can use it to compute derived functors. For the extension groups, we
get

ExtnKKG ,I (A, B) ∼
= HHn Rep(G); KK(G A, G B) ;
here HHn (R; M) denotes the nth Hochschild cohomology of a ring R with
coefficients in an R-bimodule M, and KK(G A, G B) is a bimodule over
Rep G via the Kasparov product on the left and right and the ring homomor-
phisms
Rep(G) → KK(G A, G A), Rep(G) → KK(G B, G B).
Similarly, if H : KKG → Ab commutes with direct sums, then

Ln H (B) ∼
= HHn Rep(G); H ◦ τ (G B) ,
where H ◦ τ (G B) carries the following Rep G-bimodule structure: the left

module structure comes from μB : Rep(G) → KK(G B, G B) and the
right one comes from $A : Rep(G)op → KKG (τ A, τ A) for A = G B and
the functoriality of H . The details are left to the reader.
5.1. The Pimsner–Voiculescu exact sequence

The reader may wonder why we have considered the ideal I , given that the
derived functors for I,K are so much easier to describe. This is related to the
question whether I-equivalences are invertible. The ideal I,K cannot have
this property because it already fails for trivial G. In contrast, the ideal I
sometimes has this property. This means that the spectral sequences that we
get from I may converge for all objects, not just for those in an appropriate
bootstrap category. To illustrate this, we explain how the well-known Pimsner–
Voiculescu exact sequence fits into our framework.
This exact sequence deals with actions of the group Z; to remain in the
framework of §5, we use Baaj–Skandalis duality to turn such actions into
actions of the Pontrjagin dual group T. The representation ring of T is the
ring R := Z[t, t −1 ] of Laurent polynomials or, equivalently, the group ring
of Z. An R-bimodule in an Abelian category C is an object M of C together
with two commuting automorphisms λ, ρ : M → M. The Hochschild homol-
ogy and cohomology are easy to compute using the free bimodule resolution
d
0 → R ⊗2 → R ⊗2 → R, where d(f ) = t · f · t −1 − f . We get
HH0 (R; M) ∼
= HH1 (R; M) ∼
= coker(λρ −1 − 1),
∼ HH0 (R; M) ∼
HH1 (R; M) = = ker(λρ −1 − 1),
and HHn (R; M) = ∼ HHn (R; M) = ∼ 0 for n ≥ 2. Transporting this kind of reso-
lution to C[R], we get that any object of C[R] has an I -projective resolution
of length 1. This would fail for I,K because the category of R-modules has a
non-trivial Ext2 .
The crucial point is that I -equivalences are invertible in KKT . By
Baaj–Skandalis duality, this is equivalent to the following statement: if
f ∈∈ KKZ (A, B) becomes invertible in KK, then it is already invertible
in KKZ . We do not want to discuss here how to prove this. Taking this for
granted, we can now apply Theorem 66 to all objects of KKT .
We write down the resulting exact sequences for KKZ (A, B) for
A, B ∈∈ KKZ because this equivalent setting is more familiar. The actions
of Z on A and B provide two actions of Z on KK∗ (A, B). We let
tA , tB : KK∗ (A, B) → KK∗ (A, B) be the actions of the generators. Theorem 66
yields an exact sequence

coker tA tB−1 − 1|KK∗+1 (A,B) KKZ∗ (A, B) ker tA tB−1 − 1|KK∗ (A,B) .
This is equivalent to a long exact sequence
KK1 (A, / KKZ (A, B) / KK0 (A, B)

O B) 0
tA tB−1 −1 tA tB−1 −1

KK1 (A, B) o KKZ1 (A, B) o KK0 (A, B).
Similar manipulations yield the Pimsner–Voiculescu exact sequence for the

functor A
→ K∗ (Z A) and more general functors defined on KKZ .
References
[1] Javad Asadollahi and Shokrollah Salarian, Gorenstein objects in triangulated
categories, J. Algebra 281 (2004), no. 1, 264–286. MR 2091971
[2] Saad Baaj and Georges Skandalis, C ∗ -algèbres de Hopf et théorie de Kasparov
équivariante, K-Theory 2 (1989), no. 6, 683–721 (French, with English summary).
MR 1010978
[3] Alexander A. Beı̆linson, Joseph Bernstein, and Pierre Deligne, Faisceaux pervers,
Analysis and topology on singular spaces, I (Luminy, 1981), Astérisque, vol. 100,
Soc. Math. France, Paris, 1982, pp. 5–171 (French). MR 751966
[4] Apostolos Beligiannis, Relative homological algebra and purity in triangulated

categories, J. Algebra 227 (2000), no. 1, 268–361. MR 1754234
[5] Bruce Blackadar, K-theory for operator algebras, 2nd ed., Mathematical Sciences
Research Institute Publications, vol. 5, Cambridge University Press, Cambridge,
1998. MR 1656031
[6] Alexander Bonkat, Bivariante K-Theorie für Kategorien projektiver Systeme von
C ∗ -Algebren, Ph.D. Thesis, Westf. Wilhelms-Universität Münster, 2002 (German).
electronically available at the Deutsche Nationalbibliothek at http://deposit.ddb.
de/cgi-bin/dokserv?idn=967387191.
[7] Hans-Berndt Brinkmann, Relative homological algebra and the Adams spectral
sequence, Arch. Math. (Basel) 19 (1968), 137–155. MR 0230788
[8] J. Daniel Christensen, Ideals in triangulated categories: phantoms, ghosts and
skeleta, Adv. Math. 136 (1998), no. 2, 284–339. MR 1626856
[9] Samuel Eilenberg and John Coleman Moore, Foundations of relative homological
algebra, Mem. Amer. Math. Soc. No. 55 (1965), 39. MR 0178036
[10] Peter Freyd, Representations in abelian categories, (La Jolla, Calif., 1965), Proc.
Conf. Categorical Algebra, Springer, New York, 1966, pp. 95–120. MR 0209333
[11] Pierre Julg, K-Théorie équivariante et produits croisés, C. R. Acad. Sci. Paris Sér.
I Math. 292 (1981), no. 13, 629–632 (French, with English summary).MR 625361
[12] Bernhard Keller, Derived categories and their uses, Handbook of algebra, Vol. 1,
North-Holland, Amsterdam, 1996, pp. 671–701. MR 1421815
[13] Saunders MacLane, Categories for the working mathematician, Springer-Verlag,
New York, 1971. Graduate Texts in Mathematics, Vol. 5. MR 0354798
[14] Ralf Meyer, Local and analytic cyclic homology, EMS Tracts in Mathematics,
vol. 3, European Mathematical Society (EMS), Zürich, 2007. MR 2337277
[15] , Categorical aspects of bivariant K-theory, K-theory and Noncommu-
tative Geometry (Valladolid, Spain, 2006), EMS Ser. Congr. Rep., Europ. Math.
Soc. Publ. House, Zürich, 2008, pp. 1–39.
[16] , Homological algebra in bivariant K-theory and other triangulated cate-
gories. II, Tbilisi Mathematical Journal 1 (2008), 165–210.
[17] Ralf Meyer and Ryszard Nest, The Baum–Connes conjecture via localisation of
categories, Topology 45 (2006), no. 2, 209–259. MR 2193334
[18] , An analogue of the Baum–Connes isomorphism for coactions of compact
groups, Math. Scand. 100 (2007), no. 2, 301–316. MR 2339371
[19] Amnon Neeman, Triangulated categories, Annals of Mathematics Studies,
vol. 148, Princeton University Press, Princeton, NJ, 2001. MR 1812507
[20] Radu Popescu, Coactions of Hopf-C∗ -algebras and equivariant E-theory (2004),
eprint. arXiv: math.KT/0410023.
[21] Michael Puschnigg, Diffeotopy functors of ind-algebras and local cyclic cohomol-
ogy, Doc. Math. 8 (2003), 143–245 (electronic). MR 2029166
[22] Jonathan Rosenberg and Claude Schochet, The Künneth theorem and the universal
coefficient theorem for equivariant K-theory and KK-theory, Mem. Amer. Math.
Soc. 62 (1986), no. 348. MR 0849938
[23] , The Künneth theorem and the universal coefficient theorem for Kas-
parov’s generalized K-functor, Duke Math. J. 55 (1987), no. 2, 431–474. MR
894590
[24] Ross Street, Homotopy classification of filtered complexes, J. Austral. Math. Soc.
15 (1973), 298–318. MR 0340380
[25] Jean-Louis Verdier, Des catégories dérivées des catégories abéliennes, Astérisque
(1996), no. 239, xii+253 pp. (1997) (French, with French summary). With a preface
by Luc Illusie; Edited and with a note by Georges Maltsiniotis. MR 1453167
[26] Roland Vergnioux, KK-théorie équivariante et opérateurs de Julg-Valette pour les
groupes quantiques, Ph.D. Thesis, Université Paris 7 – Denis Diderot, 2002. elec-
tronically available at http://www.math.jussieu.fr/theses/2002/vergnioux/these.
dvi.
Ralf Meyer Ryszard Nest

Mathematisches Institut and Courant Københavns Universitets Institut for
Centre “Higher order structures” Matematiske Fag
Georg-August-Universität Göttingen Universitetsparken 5
Bunsenstraße 3–5 2100 København
37073 Göttingen Denmark
Germany

Derived categories and Grothendieck duality
amnon neeman
Dedicated to Alexander Grothendieck, who started this subject, on his 80th birthday
Abstract. We study dualizing complexes. The unusual feature is that we do

not assume them to have bounded injective resolutions; we prove that the theory
works just fine with no boundedness hypothesis. In the process we prove a number
of new results about Grothendieck duality; one of the more striking is that, under
relatively mild hypotheses on f : X −→ Y , the functor f ! : D(Qcoh/Y ) −→
D(Qcoh/X) takes pseudocoherent complexes to pseudocoherent complexes. The
biggest innovation in our approach is that we systematically employ prod-
ucts in the category D(Qcoh/X); the older treatments never ventured beyond
coproducts.
In an appendix we include a proof that, if T is a stable homotopy category
in the sense of [19] in which the compact objects coincide with the strongly
dualizable objects, then the compact objects can also be characterized as those
objects such that tensoring with them respects products.
Contents
0. Introduction 291
1. Historical overview 294
2. Background on RHom complexes 301
3. Dualizing complexes 313
4. When Rf∗ respects compacts 325
5. Where we can prove Conjecture 4.16 337
6. Dualizing complexes and f ! 342
7. Several recent results 344
Appendix A. A fact concerning strongly dualizable objects 344
References 346
Key words and phrases. dualizing complex, Grothendieck duality.

The research was conducted while the author was visiting the CRM in Barcelona for six months;
many thanks to the CRM for its hospitality and congenial working environment. The work was
partly supported by a sabbatical grant from the Spanish Ministry of Education (grant number
SAB2006-0135), as well as by the Australian Research Council.
290
0. Introduction
The bulk of this survey will give a modern account of dualizing complexes. Let
X be a noetherian, separated scheme. Let Db (Coh/X) be the bounded derived
category of coherent sheaves on X. We define a dualizing complex to be an
object I ∈ Db (Coh/X) for which the functor RHom(−, I) yields an equivalence
RHom(−, I) : Db (Coh/X) −−−−→ Db (Coh/X) .

op
Note that we make no assumption that I have a bounded injective resolution;1

this is supposed to be a contemporary treatment of the subject, and by now we
should know how to handle unbounded complexes. The thrust is that the theory
works just fine. Let me now sketch the main new results which can be found in
the article.
Facts 0.1. We prove, with no simplifying hypotheses on I:
(i) A dualizing complex I, if it exists, is unique up to the obvious modifica-

tions. That is, if I and I are two dualizing complexes, then there exists
a complex L ∈ Db (Coh/X) with I ∼ = L L ⊗ I. Furthermore, on any con-
nected component of X the complex L is just a suspension of a line bundle.
See Lemma 3.9 for the proof.
(ii) An object I ∈ Db (Coh/X) is a dualizing complex if and only if it satisfies
the two conditions
(a) The natural map OX −→ RHom(I, I) is an isomorphism.
(b) For every object E ∈ Db (Coh/X) we have RHom(E, I) ∈ Db (Coh/X).
For a proof see Proposition 3.6.
The traditional reason why people cared about dualizing complexes was the
way they behave in the relative situation, where two schemes are involved; we
will discuss this a little more fully in §6. For now we confine ourselves to stating
our relevant new results. Suppose therefore that we are given a morphism f :
X −→ Y , where X and Y are noetherian, separated schemes. Let D(Qcoh/X)
and D(Qcoh/Y ) be the unbounded derived categories of quasicoherent sheaves
on X and Y respectively. There is a pushforward functor Rf∗ : D(Qcoh/X) −→
D(Qcoh/Y ), and it has a left adjoint Lf ∗ and a right adjoint f ! . The next results
are
1 The usual definition of dualizing complexes imposes the hypothesis that I have finite injective
dimension. There is a theorem of Gabber which says that the objects we call “dualizing
complexes” agree with what, in the classical literature, would go under the name “strongly
pointwise dualizing complexes”; see [6, Lemma 3.1.5]. In the classical language, one of the
things we do in this article is show that these complexes have all the good global properties of
the traditional dualizing complexes.
292 Amnon Neeman
Facts 0.2. Let I be a dualizing complex in Db (Coh/Y ) ⊂ D(Qcoh/Y ). Then

the following is true:
(i) If f is an open immersion, then Lf ∗ I is a dualizing complex in

Db (Coh/X) ⊂ D(Qcoh/X). The proof is in Remark 3.12.
(ii) Suppose f : X −→ Y satisfies the following two hypotheses:
(a) Rf∗ takes Db (Coh/X) to Db (Coh/Y ).
(b) f ! takes Db (Coh/Y ) to D+
Coh (Qcoh/X).
Then f ! I is a dualizing complex on X. The proof may be found in Theo-
rem 3.15.
Fact 0.2(ii) is a little unsatisfactory, particularly because of the hypothesis

(b) which involves the mysterious functor f ! . The next fact helps.
Facts 0.3. Hypotheses (a) and (b) of Fact 0.2(ii) hold in either of the following
two situations:
(i) The morphism f : X −→ Y is finite. Hypothesis (a) holds obviously, and

the proof that hypothesis (b) is satisfied may be found in Lemma 3.19.
(ii) The functor Rf∗ : D(Qcoh/X) −→ D(Qcoh/Y ) takes perfect complexes
to perfect complexes, and X satisfies the technical condition (∗) of Conjec-
ture 4.16. Hypothesis (a) holds by [46, Corollary 4.3.2], and for hypothesis
(b) see Corollary 0.7.
f g
As always if X −→ Y −→ Z are two morphisms, and if both f ! and g ! take
dualizing complexes to dualizing complexes, then so does the composite. We
can freely combine (i) and (ii) to produce morphisms h for which h! respects
dualizing complexes.
Remark 0.4. I do not, at the present time, fully understand the technical
condition (∗) in Fact 0.3(ii). Conjecture 4.16 articulates the hope that (∗) holds
for all X. We rephrase this slightly: assuming that the technical Conjecture 4.16
holds for every X, then f ! takes dualizing complexes to dualizing complexes as
long as Rf∗ : D(Qcoh/X) −→ D(Qcoh/Y ) takes perfect complexes to perfect
complexes.
Given that at present we do not have Conjecture 4.16 in this generality, in
addition to the preservation of perfect complexes we currently need to assume
that Conjecture 4.16 holds for X. In the remainder of the introduction I will
discuss the status of Conjecture 4.16; there are many Xs for which it is known.
Until now all our schemes were assumed noetherian. Dualizing complexes
are traditionally about producing equivalences Db (Coh/X) ∼
op
= Db (Coh/X),
and the category Db (Coh/X) does not obviously make sense unless X is noethe-
rian, or at least coherent. But it turns out that Conjecture 4.16, as well as the
technical condition (∗) in Fact 0.3(ii), make sense much more generally. For the
next result we only assume that our schemes are quasicompact and separated.
In the article we will prove the following assertion
Facts 0.5. Let f : X −→ Y be a morphism of quasicompact, separated

schemes. Assume that Conjecture 4.16 holds for X, and that Rf∗ takes perfect
complexes to perfect complexes. Then the complex f ! OY is pseudocoherent;
this means that, for any open affine subset U ⊂ X, the restriction of f ! OY
to U = Spec(R) is quasi-isomorphic to the sheafification of a bounded above
chain complex of finitely generated, projective R–modules.
The proof is in Lemma 4.21; see also Remark 4.8.
Remark 0.6. By [51, Theorem 5.4] we know that f ! S ∼ = f ! OY L ⊗ Lf ∗ S. If

∗
S is pseudocoherent then so is Lf S (obviously), and the formula, coupled
with the pseudocoherence of f ! OY of Fact 0.5, informs us that f ! S is also
pseudocoherent.
Assume now that X and Y are noetherian and S belongs to Db (Coh/Y ).
Then f ! S is pseudocoherent, which for noetherian schemes simply means it
belongs to D− (Coh/X). The fact that its cohomology is bounded below is easy
(see Remark 3.14) and we conclude the following:
Corollary 0.7. Assume f : X −→ Y is a morphism of noetherian, separated

schemes. Suppose further that Rf∗ takes perfect complexes to perfect com-
plexes, and that Conjecture 4.16 holds for X. Then f ! takes Db (Coh/Y ) to
Db (Coh/X).
In Remark 0.6 we sketched the argument leading from the more general
Fact 0.5 to Corollary 0.7, valid in the case of noetherian schemes. The reader
should note that Fact 0.3(ii) is immediate from Corollary 0.7. Corollary 0.7
proves more than we need; in Fact 0.3(ii) we only asserted that f ! takes
Db (Coh/Y ) to D+
Coh (Qcoh/X).
Facts 0.8. It remains to review what we know about Conjecture 4.16. The
current state of knowledge is that X satisfies the conjecture if
(i) X is noetherian, finite dimensional, and smooth over a finite dimensional,

noetherian regular ring R. See Theorem 5.13 for the proof.
(ii) X is a locally closed subscheme of Y , and Conjecture 4.16 is true for Y .
(iii) There is an affine morphism X −→ Y , and Conjecture 4.16 is true for Y .
Facts (ii) and (iii) both follow from Remark 5.12.
294 Amnon Neeman
We leave it to the imagination of the reader to combine these to produce many

Xs for which the conjecture is true.
This finishes the main new results we will present; it remains to give a
brief sketch of the structure of the article. We begin with a fairly extensive
historical review of Grothendieck duality. Then we cover the basic properties
of RHom(−, −); nothing here is especially new, but our approach is to define
RHom(E, −) as the right adjoint of the functor − L ⊗ E, by appealing to Brown
representability. It is therefore a little interesting to see that the theory can
be developed quickly and painlessly; this approach is relatively new, it first
appeared in Murfet’s thesis [48, Appendix C]. Then the rest of the article is
devoted to proving the claims we have made in the Introduction. The one
comment that might be helpful is that our approach hinges on systematically
using products in the categories D(Qcoh/X). These categories are compactly
generated, and therefore have products; see [52, Proposition 8.4.6]. Any right
adjoint will preserve these products, and the functors RHom(E, −), Rf∗ and
f ! are all right adjoints. The idea is to exploit this.
If X is arbitrary then products in the category D(Qcoh/X) are disgusting;
we only really understand them when X is affine. As we will see, this is often
enough. One can frequently reduce oneself to the affine case.
Acknowledgements. The author would like to thank Avramov, Iyengar, Joyal,

Krause, Lipman, Murfet, Nayak, Street and Van den Bergh for their valuable
suggestions towards improving an earlier draft. Lipman sent me a long list of
helpful comments, and Avramov and Iyengar even went as far as providing a
simplified, streamlined proof for Lemma 2.10; the current manuscript contains
their proof.
1. Historical overview
Triangulated categories came into being in the early 1960s; they arose inde-
pendently, and more or less simultaneously, in the work of Puppe [55] and of
Verdier [67]2 . Puppe’s interest in the subject came from homotopy theory; we
will say more, much later in this historical account, about the role homotopy
theory played in the early development of the subject. Our survey of the early
history will follow in the footsteps of Verdier; we will explain the problem
that inspired him, and the progress that followed. Verdier came to the subject
2 The reader should note that Verdier’s thesis was only published posthumously, many years after
it was written; the publication date of [67] is misleading.
with a specific goal in mind: he wanted to develop the necessary homological

machinery to state and prove Grothendieck duality.
Let us permit ourselves a digression, jumping ahead many years in time:
since the early days much has happened, and derived categories, or more gener-
ally triangulated categories, have successfully invaded branches of mathematics
as remote as mathematical physics [9, 18, 30, 31] and even C∗ –algebras [47];
the other articles in this volume are testament to the astounding success of
the field. My aims are more modest; in this survey we will keep our focus on
Grothendieck duality. What we will illustrate is the way in which the gradual
advances in our understanding of the foundations of derived categories, over
many years, have translated into improvements in the results on Grothendieck
duality.
Let us return to the dawn of the field; in the beginning there was the Serre
duality theorem [60]. In its original form the theorem states the following.
Theorem 1.1. Serre [60]. Let X be a connected, n–dimensional compact

complex manifold, and let V be a holomorphic vector bundle on X. Thenthe
dual of the vector space H i (X, V) is isomorphic to H n−i X , Hom(V, #n ) .
Remark 1.2. Perhaps we should explain the notation. The line bundle #n is
the bundle of holomorphic n–forms on X, and Hom(V, #n ) is the vector bundle
of whose sections, on an open set U ⊂ X, are the holomorphic maps of vector
bundles V|U −→ #n |U . There is an obvious pairing
ϕ
H i (X, V) ⊗ H n−i X , Hom(V, #n ) −−−−→ H n (#n ),
and Serre’s theorem asserts two things:
(i) The vector space H n (#n ) is one dimensional.

(ii) The pairing ϕ is perfect; it gives a natural identification of
H n−i X , Hom(V, #n ) with the dual of H i (X, V).
The question that interested people at the time was whether there is a relative
version. Suppose we are given a holomorphic map of complex manifolds f :
X −→ Y . Is there a reasonable general theorem which, in the special case where
Y = {∗} is the one-point space, comes down to Theorem 1.1? The answer turns
out to be Yes, at least in the algebro-geometric framework. In keeping with
tradition I will now switch from the complex analytic setting to the world of
algebraic geometry; I do this because complex analysis has technical difficulties
which I do not want to address.
Let X be a smooth, n–dimensional projective variety over a field k, and let
V be an algebraic vector bundle on X. Serre duality, in its algebraic variety
296 Amnon Neeman
context, gives a natural isomorphism

Hom H i (X, V) , k) −−−−→ H n−i X , Hom(V, #n ) .
To make life a little more exciting let us choose W , a finite dimensional vector
space over the field k. Clearly we also have a natural isomorphism

Hom H i (X, V) , W ) / H n−i X , Hom(V , W ⊗ #n )
k
Extn−i (V , W ⊗k #n ).
Now the vector space W can be thought of as a vector bundle over the one-point
space Y = {∗}, and, if you squint hard enough, the isomorphism above begins
to look like an adjunction. We have a functor Rf∗ , which takes a vector bundle
V on X to a string of vector bundles on Y , namely the H i (X, V). And we have
the functor f ! taking the vector bundle W on Y to W ⊗k #n . If we are willing
to treat the Ext’s as Hom’s and to disregard the confusing indices i and (n − i),
it looks like we have an isomorphism
Hom(Rf∗ V, W ) −−−−→ Hom(V, f ! W ) .

The problem was to make this intuition precise. Before all else, it would be
necessary to say exactly what Vs are allowed as inputs for the functor Rf∗ ,
and what W s as inputs for the functor f ! . We need to come up with two
categories. One, which we will provisionally label D(X), should consist of
some sort of sheaves over X. The other, for which our tentative name will be
D(Y ), will be made up of something like sheaves over Y . And we want functors
Rf∗ : D(X) −→ D(Y ) and f ! : D(Y ) −→ D(X), with f ! right adjoint to Rf∗ .
And finally we want all of this to generalize Serre duality. The objects of the
category D(X) should include the vector bundles on X, and Rf∗ V should be
something
, i which encapsulates
- the information in the string of vector spaces
H (X, V), 0 ≤ i ≤ n .
In his 1958 talk at the Edinburgh ICM (see [13]) Grothendieck announced
that he had a solution, but also that the homological algebra language necessary
to state it did not yet exist. Verdier’s thesis project was to develop the framework.
Derived categories were born in that thesis. The idea that worked, as we now
know, was to let D(X) and D(Y ) be derived categories; the objects are chain
complexes of sheaves on X and Y respectively, and the morphisms are the chain
maps, with the homology isomorphisms formally inverted.
Remark 1.3. Let us remind the reader. If f : X −→ Y is a morphism of
schemes, then there is always an induced map f∗ , taking sheaves on X to sheaves
on Y . It extends to complexes of sheaves: given a complex V of sheaves on X,

we can form a complex f∗ V of sheaves on Y . The complex Rf∗ V is obtained
by first replacing V by an injective resolution, and then applying f∗ . It is, on the
face of it, quite strange that this functor should have a right adjoint. The reason
this is a little bizarre is that the functor f∗ is left exact, but decidedly not right
exact. We would expect right adjoints to exist only for right exact functors, and
therefore f∗ most certainly cannot have a right adjoint. Only at the level of the
derived categories, after replacing f∗ by Rf∗ , do we have a hope of finding an
f ! . The passage, from the abelian category of chain complexes to its derived
category, is very brutal; it destroys exactness. We will return to this point in
Remarks 1.6 and 1.7. For now we note only that the brutality is vital, without
it Grothendieck duality wouldn’t stand a chance.
Remark 1.4. Right from the start people disliked derived categories. Derived
categories were quite unlike the more familiar objects of homological algebra,
their behaviour was poorly understood, and people felt uncomfortable using
them. The consensus was that there had to be a better, more natural substitute
for them. People felt that derived categories couldn’t possibly be the real answer,
they couldn’t be the right framework for all these theorems. To this day the
attitude persists; there are many people who are still trying, today, to find a more
natural foundation for this branch of homological algebra. A huge amount of
exciting work has come out in the last few years. The potential replacements
for derived categories include DG–categories, A∞ –categories, stable model
categories, Segal spaces, quasicategories and triangulated derivators.
Because people never liked the formalism, derived categories were slow to
catch on. Now, almost fifty years later, we can point to the impressive theorems
that have been proved using them, and to the extent to which they have come to
permeate far-flung branches of mathematics. As I have already said, the other
chapters in this book contain a compelling case for their success. At this point
we can safely say that, for a theory that everyone badmouthed from the very
outset, derived categories have come a long way.
In Remark 1.4 we mentioned that people have always had a distaste for
derived categories. It might help if we explain one of the features that no one
likes. We begin with
Lemma 1.5. Let T be a triangulated category. Then any epimorphism f :

X −→ Y in T is split.
Proof. Complete f to a distinguished triangle

f g h
X −−−−→ Y −−−−→ Z −−−−→ X.
298 Amnon Neeman
We know that the composite gf : X −→ Z vanishes, and that f is an epimor-

phism. Hence g must vanish. From [52, Corollary 1.2.7] it follows that the
triangle is isomorphic to
(1,0) 0
Y ⊕ −1 Z −−−−→ Y −−−−→ Z −−−−→ Y ⊕ Z,
and hence f : X −→ Y must be a split epimorphism.
Remark 1.6. Suppose we have a morphism f : X −→ Y , and suppose it has a
cokernel g : Y −→ Q. We observe first that the map g must be an epimorphism.
Lemma 1.5 guarantees that it is a split epimorphism: Y is isomorphic to I ⊕ Q,
where I is the (split) kernel of g. The map f : X −→ Y therefore identifies as
a composite
α
X −−−−→ I −−−−→ I ⊕ Q.
The fact that the projection I ⊕ Q −→ Q is the cokernel of f guarantees that
α : X −→ I must be an epimorphism. Using Lemma 1.5 again, we conclude
that α must split. We have shown that, if a morphism f : X −→ Y has a
cokernel, then it must factor as
α β
X −−−−→ I −−−−→ Y,
with α a split epimorphism and β a split monomorphism. In homological
algebra people had spent half a century taking kernels and cokernels of maps,
and suddenly they were faced with the world of triangulated categories, where
such constructions are worthless. No wonder they were unhappy.
Remark 1.7. Let us look at a slightly fancier version of this, where we con-
sider other possible colimits in triangulated categories. It is easy to see, by
example, that there is nothing to prevent the existence of coproducts; there
are many examples of triangulated categories with coproducts, for instance
the unbounded derived category of modules over a ring R. We say that such
categories satisfy [TR5].
Now let T be a [TR5] triangulated category. If F : S −→ T is a functor from
a small category S to T, we can wonder whether it has a colimit. Suppose such
a colimit exists. Now observe that the natural map
&
F (s) −−−−→ colim F
−→
s∈S
must be an epimorphism. Lemma 1.5 guarantees that this epimorphism has

to split. We conclude that, aside from coproducts, there can be no interesting
colimits in T. The special case of cokernels is only a baby example of a much
more serious problem.
In Remark 1.4 we mentioned that derived categories were not exactly

a welcome introduction to homological algebra, while in Lemma 1.5 and
Remarks 1.6 and 1.7 we explained one of the features that people have found
undesirable. The Almighty Lord, in His infinite grace and wisdom, has not seen
fit to bestow upon me the gift of clairvoyant, prophetic powers; it is not for
me to predict if someone really will succeed in finding a superior replacement,
whether it will be one of the contenders currently being promoted, and if so
which one. Only time will tell.
Back to Grothendieck duality. We have recounted how derived categories
were born, to establish a context in which Grothendieck’s duality theorem could
be stated and proved. The proof appeared in Hartshorne [15]; the book dates
back to 1966. Why am I wasting the reader’s time with a 40-year-old theorem?
Surely everything about it must be completely understood by now. The volume
in front of you is meant to deal with current and future mathematics, not with
ancient history.
Well, Hartshorne’s book may have appeared in 1966, but back then derived
categories were barely out of diapers, their coordination was less than perfect,
and they were still throwing temper tantrums whenever anyone tried to take
them into new territory. They were very new, poorly understood, and the tech-
nical apparatus created for dealing with them was crude and clumsy. We have
come a long way since then, and probably have an equally long road ahead of
us. There is little doubt that, forty years from now, people looking back at the
work we do today will find it laughably awkward and roundabout. We cannot
escape the fact that derived categories are work-in-progress, and many tools
are presumably yet to be conceived. This being the case we are still learning
new facts about Grothendieck duality. The main object of this article is to give
an updated account of dualizing complexes.
This is true, most of the article will be modern, but it would be wrong not
to mention the valuable contributions that people have made over the decades.
In the early years there was a great deal of interest in the subject; the theorem
was new and hot and mathematicians, like the rest of humanity, throng around
the newest hot topic on the block. Work from that period includes (of course)
Hartshorne [15], but also Deligne [7, 8], Grothendieck [14], Illusie [23, 24, 25,
26], Ramis, Ruget and Verdier [56] and Verdier [64, 65, 66].
After the early explosion of interest the subject had a long hiatus. There was
a period of almost two decades when only a small group of committed special-
ists kept the flame burning: results form this period include El Zein [11],
Grivel [12], Hopkins [16], Hopkins-Lipman [17], Hübl-Sastry [22], Hübl-
Kunz [20, 21], Kunz [36, 37], Kunz-Waldi [39], Kiehl [34], Lipman [41, 42],
Van den Bergh [63] and Yekutieli [68, 69]. But it was only in the late 1980’s that
300 Amnon Neeman
the foundations of derived categories began undergoing a major change, with

a concommitant improvement in our understanding of Grothendieck duality.
In the early days, people working on derived categories preferred to consider
their bounded incarnations. It was customary to work with D+ (A), D− (A) or
Db (A); we remind the reader that these are, respectively, the bounded below,
bounded above or bounded derived categories. Depending on the problem at
hand people were quite willing to switch around among the three bounded
versions, but the one derived category that was only rarely touched was the
unbounded derived category D(A). There was good reason for this: except in
very special cases it was not understood how to form projective or injective
resolutions of unbounded complexes, and hence there was no clear machinery
for producing derived functors. Perhaps the first article to alter matters was
Spaltenstein [61]. This 1988 paper is very methodical; it carefully outlines how
to go about this, explaining in some detail the various resolutions one can form,
and discussing the situations to which they are best suited. The terminology we
use today is still the one Spaltenstein introduced, for example when we speak
about K–injective resolutions.
The second shift came a little later, with [4, 50, 51]. It began as a pure
accident; I happened to be visiting Bielefeld, as a fellow of the Humboldt
Stiftung, in the academic year 1989–90. My office was just down the corridor
from that of Marcel Bökstedt, who is a wonderful mathematician, a clear thinker
and an excellent expositor, a man with the gift of explaining the key idea in
what would otherwise be a foggy, abstruse formalism. I enjoyed talking to him
and learned a great deal from the interaction, and in particular I learned from
Bökstedt what little homotopy theory I know. As he was explaining this to me
I was struck by the parallel with the mathematics I knew, the mathematics of
algebraic geometry. Our joint article [4] arose out of these conversations; it
amounts to my translation, to the familiar context of algebraic geometry, of the
homotopy theory that Bökstedt was teaching me.
Maybe we should elaborate a little. In Remarks 1.6 and 1.7 we noted that
there are few interesting colimits in derived categories. The homotopy theorists
have never let this bother them; they freely lift problems from triangulated cat-
egories to model categories. In the model categories colimit constructions work
just fine, and occasionally one can show that the output of these constructions
is independent of the lifting; one has produced something well-defined in the
triangulated category. It is often not difficult to reformulate the constructions in
a way that dispenses with the lifting to models. Anyway, all we really want to
stress is that the homotopy theorists walked fearlessly through terrain which the
derived categorists carefully skirted. I learned from Bökstedt that there were
paths in this unfamiliar territory, and the reader may find this in [4]. When
we worked on [4], Bökstedt and I were unaware of Spaltenstein’s earlier [61],

and we thought we were being very clever when we applied the techniques of
homotopy theory to show how one can easily handle unbounded complexes.
Unfortunately for us Spaltenstein had achieved the same, a few years earlier,
and without using any ideas from homotopy theory.
In the next few years I came to better appreciate the power of the methods,
and found a couple of new applications [50, 51]3 , results which did not tread in
the footsteps of Spaltenstein. The techniques then caught on with other authors,
and not only in the context of Grothendieck duality; for example one of the early
converts was Bernhard Keller, whose article [32] proved very influential. Some-
what later, Hovey, Palmieri and Strickland’s monograph [19] helped by mak-
ing the foundations more accessible to the non-homotopy-theorists. This survey
confines itself to Grothendieck duality; in that field, the papers which most suc-
cessfully employed homotopy-theoretic methods were Alonso, Jeremı́as and
Lipman [1, 2] Alonso, Jeremı́as and Souto [3], Jørgensen [28] and Lipman [44].
It should be mentioned that our understanding of what might be possible, using
homotopy theoretic methods and other techniques, is not yet complete; there is
ongoing work, for example [6, 38, 43, 45, 46, 49, 58, 59, 70, 71].
This ends our glimpse of history, more precisely the brief overview of the
early history of Grothendieck duality. Next we begin talking mathematics; first
we will review for the reader the basics of dualizing complexes.
2. Background on RHom complexes

Let us quickly review some basic facts about the internal Hom in the derived
category D(Qcoh/X). The results are not far from standard, but perhaps the
perspective is slightly unusual: we construct RHom using Brown representabil-
ity, and then obtain all the properties we need by reducing to the affine case.
Throughout this section X will be a quasicompact, separated scheme, and much
of the time we will also assume it noetherian. The category D(Qcoh/X) will
be the unbounded derived category of quasicoherent sheaves on X.
Remark 2.1. Instead of D(Qcoh/X) we could look at DQcoh (X), the derived
category whose objects are complexes of sheaves of OX –modules with qua-
sicoherent cohomology. When X is separated and quasicompact, the obvious
3 Once again the publication dates are misleading; the later paper [50] was published before [4],
which took longer to be accepted. With [50] I happened to be lucky; Thomason was the referee,
liked the result and accepted it immediately. Also, [50] and [51] were written at about the same
time, but published four year apart; [51] took forever to be accepted.
302 Amnon Neeman
functor D(Qcoh/X) −→ DQcoh (X) is an equivalence of categories, but for more

general X the categories can be different. For schemes that are only quasisepa-
rated it turns out that the category DQcoh (X) is the one with better properties. If
we are willing to replace D(Qcoh/X) by DQcoh (X) in what follows then many
of the results, quite possibly all, undoubtedly generalize to schemes which are
only quasiseparated. But I have not carefully checked the details.
When X is noetherian we will also consider the category Db (Coh/X); this

is our notation for the derived category of bounded complexes of coherent
sheaves. Note that coherent sheaves make sense on a noetherian scheme. The
natural functor Db (Coh/X) −→ D(Qcoh/X) is known to be fully faithful. The
essential image is the category which is sometimes denoted DbCoh (Qcoh/X);
the objects are (unbounded) complexes of quasicoherent sheaves, whose
cohomology is bounded and coherent. Since the categories Db (Coh/X) and
DbCoh (Qcoh/X) are equivalent we will freely confuse them, and use the shorter
symbol Db (Coh/X) to stand for both.
Given two objects E and F in D(Qcoh/X), we may form the derived tensor
product E L ⊗ F. It is easy to show that the derived tensor is a triangulated functor
in each of the variables E and F, and respects coproducts in either variable. If
we fix F and consider the functor E
→ E L ⊗ F, we have a triangulated functor
respecting coproducts, which we write
− L ⊗ F : D(Qcoh/X) −−−−→ D(Qcoh/X) .
We know that the category D(Qcoh/X) is compactly generated; see [51, Propo-
sition 2.5]. From [51, Theorem 4.1] it follows that the coproduct-preserving
triangulated functor − L ⊗ F must have a right adjoint; there is a functor
RHom(F, −) : D(Qcoh/X) −→ D(Qcoh/X), with

Hom E L ⊗ F , G ∼= Hom E , RHom(F, G) .
The category D(Qcoh/X) becomes a symmetric, monoidal category, and both

the tensor product and the internal Hom are triangulated functors in each of the
two variables4 .
Reminder 2.2. Next we remind the reader of a technical observation which may
be found in [48, Lemma 6.8]. For this observation we need a little more notation.
4 It is slightly subtle to show that the internal Hom is triangulated in the first variable. For details
see Murfet’s thesis [48, Theorem C.1]. There are also older proofs in the literature, for example
[44, §1.5.3]. Unlike the older arguments, the proof in [48] is in the spirit of this section; it
appeals to the definition of the internal Hom as right adjoint to the tensor product, and develops
the property of the tensor product that would suffice to formally deduce that the right adjoint
must be triangulated in the first variable.
The scheme X is still assumed quasicompact and separated, but now we wish
to also consider a quasicompact open subset U ⊂ X. We will let j : U −→ X
stand for the inclusion. Because we now have two schemes, we distinguish
RHomU (F, G) from RHomX (F, G); if F, G are objects of D(Qcoh/U ) then the
first is the RHom which makes sense, while if F, G belong to D(Qcoh/X), then
the second is well-defined.
The remark we want is that, if F ∈ D(Qcoh/X) and G ∈ D(Qcoh/U ), then
there is a natural isomorphism
RHomX (F, Rj∗ G) ∼
= Rj∗ RHomU (j ∗ F, G).
The proof is easy; the result follows from the isomorphisms

HomX E , RHomX (F, Rj∗ G) ∼ = HomX E L ⊗ F , Rj∗ G

∼
= HomU j ∗ E L ⊗ j ∗ F , G

∼
= HomU j ∗ E , RHomU (j ∗ F, G)

∼
= Hom E , Rj∗ RHom (j ∗ F, G) .
X U
Remark 2.3. With the notation as in Reminder 2.2 we observe that,

for any pair of objects F, G ∈ D(Qcoh/X), there is always a map
γ (F, G) : j ∗ RHomX (F, G) −→ RHomU (j ∗ F, j ∗ G). To construct it, consider
the sequence of maps

HomX E , RHomX (F, G) HomX E L ⊗ F , G
ggg
ggggggggg
gggg
∗ L sggggg
HomU j E ⊗ j ∗ F , j ∗ G HomU j ∗ E , RHomU (j ∗ F, j ∗ G)

HomX E , Rj∗ RHomU (j ∗ F, j ∗ G)
They are all natural in E and we deduce, in the category D(Qcoh/X), a

morphism RHomX (F, G) −→ Rj∗ RHomU (j ∗ F, j ∗ G). Adjunction gives us, in
the category D(Qcoh/U ), the required map γ (F, G) : j ∗ RHomX (F, G) −→
RHomU (j ∗ F, j ∗ G).
The morphism γ (F, G) need not be an isomorphism. In general, all we know
for certain is that it is natural in F and G. There is, however, a useful special
case in which we can show γ (F, G) to be an isomorphism.
Lemma 2.4. Assume that U ⊂ X is a quasicompact open subset of a quasi-
compact, separated scheme X. Suppose G ∈ D(Qcoh/X) is any object, and let
F ∈ D(Qcoh/X) satisfy one of the two possible conditions:
304 Amnon Neeman
(i) X is a noetherian scheme, and F belongs to Db (Coh/X) ⊂ D(Qcoh/X).

(ii) X is general, but we assume F is a compact object in D(Qcoh/X). We
remind the reader: the compact objects are the perfect complexes.
If either (i) or (ii) holds, then the morphism γ (F, G) : j ∗ RHomX (F, G) −→
RHomU (j ∗ F, j ∗ G) of Remark 2.3 is an isomorphism.
Proof. Assume first that both U and X are affine. In that case we have rings
R and S so that X = Spec(R) and U = Spec(S), and we know that the open
immersion j : U −→ X is flat, meaning that S must be a flat R–algebra. The
derived category D(Qcoh/X) identifies with D(R–Mod), the derived category
D(Qcoh/U ) is D(S–Mod), and the functor j ∗ is nothing more than tensoring a
chain complex of R–modules with S. The lemma is now an easy consequence
of the observation that, if M and N are R–modules with M finitely presented,
then the S–module homomorphism
S ⊗R HomR (M, N) −−−−→ HomR (S ⊗R M , S ⊗R N )
is an isomorphism.
Next we will see how to deduce the case where only U is assumed affine.
We note that, if we fix F and let G vary, the maps γ (F, G) assemble to a natural
transformation between triangulated functors. To show that the maps γ (F, G)
are isomorphisms, for fixed F and all G, it therefore suffices to prove the special
case where G belongs to some class of generators for D(Qcoh/X). The class
we will use is the objects {jVX }∗ G = R{jVX }∗ G, where jVX : V −→ X is the
open immersion of an open affine subset V ⊂ X. Observe that, since V ⊂ X is
assumed affine, the functor {jVX }∗ is exact; hence {jVX }∗ G = R{jVX }∗ G. See [48,
Corollary 3.14] for the fact that these generate.
Because we now have several open subsets involved, we will adopt the
convention that jWW12 : W1 −→ W2 stands for the open immersion between two
subsets W1 ⊂ W2 of X.
The lemma now follows from the isomorphisms
∗ ∗ ∗
{jUX } RHom F , {jVX }∗ G ∼= {jUX } {jVX }∗ RHom {jVX } F , G
∗ ∗
∼
= {jUU∩V }∗ {jUV ∩V } RHom {jVX } F , G
∗ ∗ ∗
∼
= {jUU∩V }∗ RHom {jUV ∩V } {jVX } F , {jUV ∩V } G
∗ ∗ ∗
∼
= {jUU∩V }∗ RHom {jUU∩V } {jUX } F , {jUV ∩V } G
∗ ∗
∼
= RHom {jUX } F , {jUU∩V }∗ {jUV ∩V } G
∗ ∗
∼
= RHom {j X } F , {j X } {j X } G .
U U V ∗
The first isomorphism is by Reminder 2.2, the second is by base change, the
third is by the affine case which we already know, the fourth is trivial, the fifth
is again by Reminder 2.2, and finally the sixth is yet another base change.
Finally we need to see how to deduce the general case, where neither U nor
X need be affine. We wish to show that the map γ (F, G) : j ∗ RHomX (F, G) −→
RHomU (j ∗ F, j ∗ G) is an isomorphism. The problem is local; it suffices to show
that it restricts to an isomorphism on each affine open set V ⊂ U . Choose such
an affine open set, and let i : V −→ U be the inclusion. It suffices to show that
the map
i ∗ γ (F, G) : i ∗ j ∗ RHomX (F, G) −−−−→ i ∗ RHomU (j ∗ F, j ∗ G)
is an isomorphism. But we already know that the inclusions i : V −→ U and
j i : V −→ X satisfy the Lemma, since V is affine. Hence both complexes are
naturally isomorphic to RHomV (i ∗ j ∗ F, i ∗ j ∗ G).
Remark 2.5. Suppose X is a quasicompact, separated scheme. Given
two complexes E, F ∈ D(Qcoh/X) we have a counit of adjunction
ε : E L ⊗ RHom(E, F) −→ F. If we have two more complexes E , F ∈
D(Qcoh/X) they also give a map ε : E L ⊗ RHom(E , F ) −→ F . Combin-
ing the two we obtain a single morphism
ε L⊗ ε
E L ⊗ RHom(E, F) L ⊗ E L ⊗ RHom(E , F ) −−−−→ F L ⊗ F ,
and adjunction produces for us a map
RHom(E, F) L ⊗ RHom(E , F ) −−−−→ RHom(E L ⊗ E , F L ⊗ F ).
The special case where E = OX will interest us particularly; in this case we
have a morphism
μ(E, F, F ) : RHom(E, F) L ⊗ F −−−−→ RHom(E , F L ⊗ F ).
We prove:
Lemma 2.6. Let X be a quasicompact, separated scheme. If E ∈ D(Qcoh/X)
is a perfect complex, and F and F in D(Qcoh/X) are arbitrary, then the
morphism μ(E, F, F ) of Remark 2.5 is an isomorphism.
Proof. We have a globally defined morphism in D(Qcoh/X) and wish to prove
it an isomorphism; it suffices to show that, for any open affine subset U ⊂ X,
the morphism μ(E, F, F ) restricts to an isomorphism on U . Let j : U −→ X
be the inclusion. Lemma 2.4 permits us to identify
j ∗ RHom(E, F) ∼ = RHom(j ∗ E, j ∗ F) ,
j ∗ RHom(E, F L ⊗ F ) ∼
= RHom(j ∗ E , j ∗ F L ⊗ j ∗ F )
306 Amnon Neeman
and we conclude that j ∗ μ(E, F, F ) = μ(j ∗ E, j ∗ F, j ∗ F ). We are therefore

reduced to proving the Lemma in the case where X is affine. Put X = Spec(R);
then the category D(Qcoh/X) identifies with D(R–Mod), and under the identifi-
cation the perfect complex E becomes a bounded complex of finitely generated
projectives. Because μ(E, F, F ) is a natural tranformation of triangulated func-
tors in E, we easily reduce to the case where E is a single, finitely generated
free module concentrated in degree 0, and in this case the result is obvious.
Reminder 2.7. Consider the special case of Lemma 2.6, where we let F = OX .
From Lemma 2.6 we learn that, for every perfect complex E and for any object
F ∈ D(Qcoh/X), the natural map
RHom(E, OX ) L ⊗ F −−−−→ RHom(E, F )

is an isomorphism. This brings us to the well-explored realm of strongly dual-
izable objects. We remind the reader.
It is traditional to define E∨ = RHom(E, OX ). The standard terminology is
that E is strongly dualizable if the natural map E∨ L ⊗ F −→ RHom(E, F ) is
an isomorphism; the previous paragraph taught us that all perfect complexes
are strongly dualizable. From the literature on strongly dualizable objects we
learn that
(i) If E is a strongly dualizable object then the natural map E −→ {E∨ }∨ is
an isomorphism. See [40, Proposition 1.3(i), p. 122].
(ii) If E is strongly dualizable then so is E∨ . See [40, Proposition 1.2, p. 121]5 .
(iii) Let T be a compactly generated triangulated category, with a symmetric
tensor product compatible with the triangulated structure. Assume that the
unit of the tensor is compact, and that all compact objects are strongly
dualizable. For example T could be D(Qcoh/X). If E ∈ T is some object,
then the following are equivalent:
(a) E is compact.
(b) E is strongly dualizable.
(c) Tensor product with E commutes with arbitrary products in T; that is,
the natural map

E L⊗ tλ −−−−→ (E L ⊗ tλ )
λ∈ λ∈
is an isomorphism for every set of objects {tλ , λ ∈ }.
5 The objects we call “strongly dualizable” are labeled “finite” in [40]. In the second last
paragraph on [40, p. 120] the authors refer to [10] for the older terminolgy, which is the one in
current use. Similar notions appeared even earlier, under different names; see the “⊗–categories
rigides” of [57, §I.5, p. 78], or the “compact closed categories” of [33, pp. 102–103].
The fact that (c) implies (a) seems new; in Theorem A.1 we will give a
self-contained proof that (a), (b) and (c) are equivalent.
In subsequent sections, especially §4, we will feel free to use (i), (ii) and (iii)
above.
Remark 2.8. The counit of adjunction gives a natural map ε :

E L ⊗ RHom(E, F) −→ F. Combining two such, we deduce a composite
ε⊗1 ε
E L ⊗ RHom(E, F) L ⊗ RHom(F, G) −−−−→ F L ⊗ RHom(F, G) −−−−→ G .
By adjunction this corresponds to a morphism

ν(E, F, G) : E L ⊗ RHom(F, G) −−−−→ RHom RHom(E, F) , G .
We will find useful the next little lemma; note that, from now until the end of
the section, all our schemes are assumed noetherian.
Lemma 2.9. Let X be a noetherian, separated scheme. Let the notation be

as above, with the objects F, G ∈ D(Qcoh/X) fixed. Suppose that, for all E ∈
Db (Coh/X),
(i) The complex RHom(E, F) belongs to Db (Coh/X) ⊂ D(Qcoh/X). The case

E = OX tells us thatF ∈ Db (Coh/X).
(ii) The complex RHom RHom(E, F) , G belongs to D− (Qcoh/X).
Then, for all E ∈ Db (Coh/X), the map ν(E, F, G) is an isomorphism.
Proof. We have a globally defined morphism in the derived category, and we

wish to prove it a homology isomorphism. The question is local; it suffices to
show that the restriction of ν(E, F, G) to every open affine subset U ⊂ X is a
homology isomorphism. Choose any open affine U ⊂ X, and let j : U −→ X
be the inclusion. We propose to show that the map

j ∗ ν(E, F, G) : j ∗ E L ⊗ j ∗ RHom(F, G) −−−−→ j ∗ RHom RHom(E, F) , G .
is an isomorphism in D(Qcoh/U ).
At this point we use hypothesis (i) and Lemma 2.4. Because F ∈
D (Coh/X), we have that j ∗ RHom(F, G) = RHom(j ∗ F, j ∗ G), because E ∈
b
Db (Coh/X) we know that j ∗ RHom(E, F) = RHom(j ∗ E, j ∗ F), and because

RHom(E, F) ∈ Db (Coh/X) we conclude that

j ∗ RHom RHom(E, F), G = RHom j ∗ RHom(E, F), j ∗ G

= RHom RHom(j ∗ E, j ∗ F), j ∗ G .
308 Amnon Neeman
We are therefore reduced to showing that the map
ν(j ∗ E, j ∗ F, j ∗ G) : j ∗ E L ⊗ RHom(j ∗ F, j ∗ G) −→

RHom RHom(j ∗ E, j ∗ F), j ∗ G
is an isomorphism in D(Qcoh/U ); in other words it suffices to prove the lemma

in the special case where X = U is affine.
Assume therefore that X is affine. Then X = Spec(R) for a noetherian ring
R, and D(Qcoh/X) is identified with D(R–Mod). The complex E becomes a
bounded complex of finitely generated R–modules, and, because ν(E, F, G)
is a natural transformation between triangulated functors in E, it suffices to
prove ν(E, F, G) to be an isomorphism in the special case where E is a single,
finitely generated module concentrated in degree 0. Next observe that, for any
V, W ∈ D(R–Mod), we may construct RHom(V, W) by first replacing W by
a K–injective resolution consisting of injective objects (see, for example, [4,
Applications 2.4 and 2.4’]), and then computing the ordinary Hom–complex.
For the problem at hand, let us replace F and G by K–injective resolutions of
injectives. All the complexes RHom(V, W), on both sides of the map ν(E, F, G),
simplify to Hom(V, W). Note also that, because R is noetherian, we know that
Hom(F, G) is a complex of flat modules; for any two injective modules I and J
we have that Hom(I, J ) is flat, and products of flat modules are flat. However,
we do not yet know that the complex Hom(F, G) is K–flat; in other words there
remains a derived tensor product which we will eventually need to handle. We
must show that the morphism

ν(E, F, G) : E L ⊗ Hom(F, G) −−−−→ Hom Hom(E, F), G
is an isomorphism in D(R–Mod).
With our choices of complexes, that is where F and G have been replaced by
K–injective resolutions by injectives and where E is a single, finitely generated
module in degree 0, we obtain a chain map of chain complexes

α : E ⊗ Hom(F, G) −−−−→ Hom Hom(E, F), G .
Do not confuse α with the map ν(E, F, G); we are not yet asserting that they
agree. We will now prove that the morphism α is an isomorphism of chain
complexes. To see this, choose a finite presentation of the module E. That is,
choose an exact sequence
R m −−−−→ R n −−−−→ E −−−−→ 0.

We deduce a diagram of chain maps of chain complexes

R m ⊗ Hom(F, G) / Rn ⊗ Hom(F, G) / E ⊗ Hom(F, G) /0
γ β α

Hom Hom(R m , F) , G / Hom Hom(R n , F) , G / Hom Hom(E, F) , G /0
The rows are clearly exact, we know that β and

γ are isomorphisms,
and hence
so is α. We have computed the complex Hom Hom(E, F) , G , and the lemma
now reduces to showing that the natural map
E L ⊗ Hom(F, G) −−−−→ E ⊗ Hom(F, G)
is a homology isomorphism. We must show that the derived tensor product
agrees with the simple-minded tensor product. It certainly suffices to prove that
Hom(F, G) is K–flat.
Now we use Hypothesis (ii) of the Lemma. We are given that, for any finite
R–module E, the complex

E ⊗ Hom(F, G) ∼= Hom Hom(E, F), G
belongs to D− (R–Mod); its cohomology is bounded above. Lemma 2.9 there-
fore follows from the following technical lemma.
Lemma 2.10. Let R be a commutative, noetherian ring, and let H be a complex
of flat R–modules. Suppose that, for any finite R–module M, we have
(i) H i (M ⊗ H) = 0 for i 0. More precisely: for every M there exists an
integer N = N (M), which may depend on M, so that i > N (M) implies
H i (M ⊗ H) = 0.
Then the complex H is K–flat.
Remark 2.11. The proof we give for Lemma 2.10 is due to Avramov and Iyen-
gar; it is a simplified version of my original, clumsy argument. The remainder
of the section is devoted to the proof; the reader may want to skip ahead to §3.
In subsequent sections we will not make use of any of the ideas in the proof.
Proof. We need to show that H is K–flat. We remind the reader: this means
that X ⊗ H must be shown acyclic for every acyclic complex X of R–modules.
We will break up the proof into steps. We begin with
Step 1. It suffices to prove the Lemma under the extra hypotheses that H is
acyclic.
Proof. Suppose H is an object in K(R–Flat) satisfying the hypotheses of the
Lemma; H does not have to be acyclic. Let us apply the hypothesis (i) of
310 Amnon Neeman
the Lemma in the special case M = R; we have a finite module R, and (i)
tells us that H i (H) = H i (R ⊗ H) vanishes for i 0. Choose a K–projective
resolution of G −→ H. The map G −→ H is a quasi-isomorphism, and the
complex G may be chosen to belong to K− (R–Proj); that is we can take G to be
a bounded above complex of projective R–modules. Complete the morphism
G −→ H to a distinguished triangle in K(R–Flat)
G −−−−→ H −−−−→ H −−−−→ G;
this produces an acyclic complex H of flat R–modules. We furthermore know

that G, being a bounded above complex of projectives, is definitely K–flat, and
it therefore suffices to show that H is K–flat. The hypothesis (i) of the Lemma
holds for H by assumption, and for G because it is a bounded above complex.
Hence H also satisfies the hypotheses of the Lemma. Replacing H by H , we
are reduced to proving the Lemma when the object H is acyclic.
Strategy: We wish to show that X ⊗ H is acyclic when X is. It suffices to

show the acyclicity of the localizations (X ⊗ H)p = X ⊗ Hp for every prime
ideal p ⊂ R. That is, it suffices to prove that each Hp is K–flat. We will prove
this by induction on the height of the prime ideal p. To give away our strategy
even more completely, we will consider the two conditions
(ii) For every prime ideal of height ≤ n the complex Hp is K–flat.

(iii) For every prime ideal p ⊂ R of height ≤ n, for every finite R–module M,
and for every integer i, we have that H i (M ⊗ Hp ) is p–torsion.
Our induction will be to show that if (ii) is true for n then (iii) is true for n + 1,
and if (iii) is true for n then (ii) is true for n. Note that (iii) is obviously true for
n = 0; this starts the induction.
Step 2. If (ii) is true for n then (iii) is true for n + 1.
Proof. Let p be a prime ideal of height n + 1 and let M be a finite R–module.

We wish to show that the Rp –modules H i (M ⊗ Hp ) are all p–torsion; it suffices
to show that their localizations at prime ideals q vanish, as long as q is properly
contained in p. Now localization is flat, and hence
* i +
H (M ⊗ Hp ) q ∼ = H i (M ⊗ Hq ) ;
it suffices to show that each complex M ⊗ Hq is acyclic. Since q is properly

contained in p it has height ≤ n, and, because (ii) is true for n, we know that
Hq is K–flat. This makes it a K–flat, acyclic complex. The acyclicity says that,
in the derived category D(R–Mod), the complex Hq is isomorphic to zero. The
fact that it is K–flat permits us to compute, in D(R–Mod),
M ⊗ Hq ∼
= M L ⊗ Hq ∼
= M L ⊗ 0 = 0.
This gives the acyclicity of M ⊗ Hq .
Notation 2.12. It remains to prove that if (iii) is true for n then so is (ii). We fix
a prime ideal p of height n; we want to prove that Hp is K–flat. Let us simplify
the notation a little. Observe first that X ⊗R Hp ∼ = Xp ⊗Rp Hp ; replacing R
by Rp we may assume R is a local ring of height n, with maximal ideal p
and residue field k = R/p. Let us further replace H by the chain complex Hp .
What we know so far is:
(iv) For every finite R–module M we have H i (M ⊗ H) = 0 if i 0; this

comes from hypothesis (i) of the Lemma.
(v) The complex H is an acyclic complex of flat R–modules; that comes from
Step 1.
(vi) For all finite R–modules M and for all integers i we have that H i (M ⊗ H)
is p–torsion. This is because we are assuming that (iii) holds for n.
We want to prove that, under these conditions, H must be K–flat.
Step 3. With the conventions of Notation 2.12, there exists an integer so that
H i (M ⊗ H) vanishes for all finite R–modules M and for all i > .
Proof. The proof will appeal to the following observation: H is a complex of

flat modules, and therefore any short exact sequence 0 −→ M −→ M −→
M −→ 0 of R–modules induces a short exact sequence of chain complexes
0 −−−−→ M ⊗ H −−−−→ M ⊗ H −−−−→ M ⊗ H −−−−→ 0,

hence a long exact sequence in cohomology.
By (iv) we know that H i (k ⊗ H) vanishes for i 0. Shifting H if necessary,
we may suppose that H i (k ⊗ H) vanishes for i > 0. If M is a finite R–module
with dim(M) = 0, then M has a finite filtration with subquotients isomorphic
to k. From the exact sequences in cohomology of the previous paragraph we
immediately learn that H i (M ⊗ H) must also vanish whenever i > 0.
We will prove, by induction on m = dim(M), that H i (M ⊗ H) vanishes
if i > dim(M). The previous paragraph proved the case m = 0. Since every
module has dimension dim(M) ≤ dim(R) = n, we will conclude that H i (M ⊗
H) = 0 for any finite module M and all i > n; that is Step 3 will immediately
follow.
Suppose m is an integer ≥ 0, and assume we know the vanishing of H i (N ⊗
H) provided dim(N ) ≤ m and i > dim(N). Let M be a finite R–module with
312 Amnon Neeman
dim(M) = m + 1. Let p (M) be the p–torsion submodule of M. We have a

short exact sequence of R–modules
0 −−−−→ p M −−−−→ M −−−−→ M/ p M −−−−→ 0.
Since the module p M is zero-dimensional we know the vanishing of

p M ⊗ H) for all i > 0; this means that the map H (M ⊗ H) −→
H i ( i
H [M/ p M] ⊗ H is an isomorphism for i > 0. Replacing M by M/ p M

i
we may assume that M has no p–torsion. We may choose an element x ∈ p

which is not a zero-divisor on M. Consider the exact sequence
x
0 −−−−→ M −−−−→ M −−−−→ M/xM −−−−→ 0;
the long exact sequence in cohomology tells us the exactness of
x
H i [M/xM] ⊗ H −−−−→ H i+1 (M ⊗ H) −−−−→ H i+1 (M ⊗ H).
The module M/xM has dimension m, and the inductive hypothesis establishes
the vanishing of H i [M/xM] ⊗ H if i > m. The exact sequence implies
that multiplication by x ∈ p is injective on H i (M ⊗ H) when i ≥ m + 1. But
(vi) informs us that the module H i (M ⊗ H) is p–torsion, and hence must
vanish.
Step 4. With the conventions of Notation 2.12, the complex H is K–flat.
Proof. Let the chain complex H be written as
· · · −→ Hi−2 −→ Hi−1 −→ Hi −→ Hi+1 −→ Hi+2 −→ · · ·
For each i ∈ Z let I i be the image of the homomorphism Hi −→ Hi+1 . The

acyclicity of H tells us that the sequence
Hi−2 −−−−→ Hi−1 −−−−→ Hi −−−−→ I i −−−−→ 0

is exact; it is the beginning of a flat resolution for the module I i . Now let M be
any finite R–module, and consider the sequence
M ⊗ Hi−2 −−−−→ M ⊗ Hi−1 −−−−→ M ⊗ Hi −−−−→ M ⊗ I i −−−−→ 0.

In this sequence, the part
M ⊗ Hi−1 −−−−→ M ⊗ Hi −−−−→ M ⊗ I i −−−−→ 0

must be exact, simply from the right exactness of the tensor product. If i >
+ 1, with as in Step 3, then the bit
M ⊗ Hi−2 −−−−→ M ⊗ Hi−1 −−−−→ M ⊗ Hi

must also be exact, because H i−1 (M ⊗ H) = 0 for i − 1 > . We conclude

that, as long as i > + 1, the torsion group TorR1 (M, I i ) vanishes for all finite
modules M. That is, I i is flat if i > + 1.
Now consider the short exact sequences
0 −−−−→ I i−1 −−−−→ Hi −−−−→ I i −−−−→ 0.

We know that Hi is flat for all i. If I i is flat, the sequence tells us that so is
I i−1 . We know that, for sufficiently large i, the modules I i are flat, and by
descending induction we deduce the flatness of I i for every integer i. The short
exact sequence
0 −−−−→ I i−1 −−−−→ Hi −−−−→ I i −−−−→ 0

is therefore a sequence of flat modules, and hence the sequences
0 −−−−→ M ⊗ I i−1 −−−−→ M ⊗ Hi −−−−→ M ⊗ I i −−−−→ 0

are all exact. Piecing them all together, we conclude that the sequence
· · · −−−−→ M ⊗ Hi−1 −−−−→ M ⊗ Hi −−−−→ M ⊗ Hi+1 −−−−→ · · ·

is exact. That is the complex M ⊗ H is acyclic.
We now know that M ⊗ H is acyclic for any finite module M. It is standard
that X ⊗ H must then be acyclic for any chain complex X of R–modules; see,
for example, [53, Corollary 9.4, (iii)=⇒(ii)]. In particular X ⊗ H is acyclic
when X is, that is H is K–flat.
3. Dualizing complexes
Throughout this section we continue to assume, as in the second half of §2, that
X is a noetherian, separated scheme. In §2 we prepared the ground by setting up
some technical apparatus, and now we are ready to treat dualizing complexes.
Let us first remind the reader of the definition.
Definition 3.1. Suppose, as agreed, that X is a noetherian, separated scheme.

A dualizing complex for X is an object I ∈ Db (Coh/X) so that the functor
E
→ RHom(E, I)
gives an equivalence
RHom(−, I) : Db (Coh/X) −−−−→ Db (Coh/X).

op
314 Amnon Neeman
Remark 3.2. Perhaps we should remind the reader: when E and I are two
objects of Db (Coh/X), the complex RHom(E, I) has no obligation to belong
to Db (Coh/X); there is no reason to expect its cohomology to be bounded
above. It might help to recall the special case when X = Spec(R) is affine, and
Db (Coh/X) reduces to Db (R–mod). Consider two finite R–modules M and N ;
they are objects of Db (R–mod), and RHom(M, N ) is a chain complex whose
cohomology is Extn (M, N). It is perfectly possible to have Extn (M, N ) = 0
for infinitely many n.
As the above example illustrates it is a restriction on I to demand that, for
every E, the object RHom(E, I) be isomorphic to an object in Db (Coh/X). It
is a severe restriction to further insist that the functor
RHom(−, I) : Db (Coh/X) −−−−→ Db (Coh/X)
op
be an equivalence. In this section we will study some of the consequences.

Remark 3.3. The experts will notice that Definition 3.1 is unorthodox. It is
customary to impose on I the technical condition that it have a finite injective
resolution; that is, one traditionally assumes a dualizing complex I to be quasi-
isomorphic to a bounded complex of injectives.
The whole thrust of this manuscript is that by now, more than forty years after
derived categories were introduced, we have learnt enough about unbounded
complexes to be able to handle them without trembling. It would be wimpy
to assume boundedness, and we have decided to be tough. But the reader
should be warned that, as a result of our decision to take the macho approach,
Definition 3.1 is a little non-standard, as is what will now follow. What we
will do amounts to proving that the usual results can be obtained without the
boundedness hypothesis. And the reason for §2 was to provide us with the
technical lemmas we will need.
Remark 3.4. Next observe that, for any object G ∈ D(Qcoh/X), the func-
tor RHom(−, G) : D(Qcoh/X)op −→ D(Qcoh/X) is its own left adjoint; this
comes from the isomorphisms

Hom E , RHom(F, G) ∼ = Hom E L ⊗ F , G

∼
= Hom F , RHom(E, G) .
Furthermore, the unit and counit of this adjunction are the same morphism;
they are both the natural map

E −−−−→ RHom RHom(E, G) , G .
If G is carefully chosen, so that RHom(−, G) takes objects in Db (Coh/X) to
objects in Db (Coh/X), then the adjunction restricts from the large category to
the subcategory; that is
RHom(−, G) : Db (Coh/X) −−−−→ Db (Coh/X)

op
is its own left adjoint, and the unit and counit of adjunction are the restrictions
to Db (Coh/X) of the unit and counit of adjunction on D(Qcoh/X).
Now general category theory kicks in and tells us that, if G is a functor with
a left adjoint F , then G will be an equivalence if and only if both the unit and
the counit of adjunction are isomorphisms. In our particular case, where the
unit and counit happen to agree, this comes down to the following.
Lemma 3.5. Choose an object I ∈ Db (Coh/X). The object I is a dualizing

complex if and only if, for every object E ∈ Db (Coh/X), we have
(i) RHom(E, I) ∈ Db (Coh/X).

(ii) The map E −→ RHom RHom(E, I), I is an isomorphism.
Proof. (i) is equivalent to RHom(−, I) restricting to a functor from Db (Coh/X)

to itself, while (ii) is equivalent to the unit and counit of the adjunction of
Remark 3.4 being isomorphisms.
Slightly less trivial is the formulation
Proposition 3.6. Let I be an object of Db (Coh/X). Then I is a dualizing

complex if and only if
(i) For every object E ∈ Db (Coh/X) we have RHom(E, I) ∈ Db (Coh/X).

(ii) The natural map OX −→ RHom(I, I) is an isomorphism.
Proof. To deduce Proposition 3.6 from Lemma 3.5, we need to show that it
suffices to check that the morphism

E −−−−→ RHom RHom(E, I), I
is an isomorphism in the special case where E = OX . The useful Lemma 2.9

allows us to identify, for any E,

E L ⊗ RHom(I, I) ∼
= RHom RHom(E, I), I .
The morphism we need to show an isomorphism becomes
E −−−−→ E L ⊗ RHom(I, I).
We leave it to the reader to check that this morphism is 1 ⊗ ρ, where 1 : E −→ E

is the identity and ρ : OX −→ RHom(I, I) is the natural map. It therefore
suffices to show that ρ is an isomorphism.
316 Amnon Neeman
Notation 3.7. If X is a noetherian scheme and I is a dualizing complex on

X, we will sometimes abbreviate the functor RHom(−, I) to just DX,I (−). If
either X or I is understood from the context we will feel free to drop it from the
notation; thus the symbols DX (−), DI (−) and DX,I (−) should be viewed as
synonymous, where the third is the most explicit. What we constructed above,

using the unit of adjunction, is a natural isomorphism E −→ DX,I DX,I (E) .
Remark 3.8. The next obvious question is what happens if we are given,
on the same scheme X, two dualizing complexes I and J. How do they
compare?
It is clear that, if we are given a dualizing complex I, an integer n and a line
bundle L, then n L L ⊗ I is also a dualizing complex; tensoring with a line
bundle and suspending is harmless. It is also harmless to suspend by different
integers on different connected components of X. The remarkable fact is that
this is all the freedom we have. Up to these basic moves, the dualizing complex
is unique. We state this as a lemma and give the proof.
Lemma 3.9. Let I and J be two dualizing complexes on the same scheme X.
Then I = L L ⊗ J, where L is locally some suspension of a line bundle. On each
connected component Xi ⊂ X there is an integer ni , a line bundle Li , and an
isomorphism L|Xi ∼= ni Li .
Proof. Let us define L = RHom(I, J). Now note that
J = RHom(O X , J)
= RHom RHom(I, I), J by Proposition 3.6(ii)
= I L ⊗ RHom(I, J) by Lemma 2.10
= I L⊗ L by definition of L .
It remains to analyse L and prove that, on each connected component of X, it

is isomorphic to the suspension of a line bundle.
Consider the functor DI DJ . This functor takes the complex E to the complex

DI DJ (E) = RHom RHom(E, J), I
E ⊗ RHom(J, I). Interchanging

L
and Lemma 2.9 identifiesthis with the roles of
I and J, we have that DJ DI (E) is naturally isomorphic to E L ⊗ RHom(I, J).
Now the functor DI DJ DJ DI is naturally isomorphic to the identity; if the
units=counits of the adjunctions of Remark 3.4 are written ηi : 1 =⇒ DI DI
and ηj : 1 =⇒ DJ DJ , then the composite
ηi DI ηj D I
1 −−−−→ DI DI −−−−−→ DI DJ DJ DI
provides the isomorphism. On the object E ∈ Db (Coh/X) this isomorphism is

the natural map
E −−−−→ E L ⊗ RHom(I, I) −−−−→ E L ⊗ RHom(I, J) L ⊗ RHom(J, I).
Applying this to E = OX , we see that
RHom(I, J) L ⊗ RHom(J, I) ∼
= OX .
We already defined L = RHom(I, J); now set L = RHom(J, I), and we
deduce an isomorphism L L ⊗ L ∼ = OX . We furthermore know that L and L
b
are both objects in D (Coh/X).
It remains to show that an object L ∈ Db (Coh/X), which has an inverse L
with respect to the tensor product, must locally be a shift of a line bundle; we
remind the reader of the proof. Restrict everything to the stalk at a point x ∈ X.
Then the stalk at x of the sheaf OX is a local ring R with maximal ideal m.
Let us write Lx , Lx for the stalks at x of L, L . Then Lx , Lx are objects in
Db (R–mod) satisfying Lx L ⊗ Lx ∼ = R. Let k = k(x) be the residue field of R.
We know that k L ⊗ Lx is a complex of k–vector spaces; that is it must be a
direct sum of suspensions of k. Now consider the isomorphisms
(k L ⊗ Lx ) L ⊗ Lx = k L ⊗ (Lx L ⊗ Lx ) = k L ⊗ R = k .
We learn first that k L ⊗ Lx must be nonzero, and by symmetry k L ⊗ Lx is
also nonzero. But also k = (k L ⊗ Lx ) L ⊗ Lx is indecomposable. It follows
that k L ⊗ Lx cannot be a direct sum of more than one factor, since then
(k L ⊗ Lx ) L ⊗ Lx would decompose into more than one direct summand, and
each of the summands would be isomorphic to some suspension of the non-
trivial k L ⊗ Lx . We conclude that k L ⊗ Lx must be of the form n k, for some
integer n.
Now Lx is an object in Db (R–mod), and it therefore has a minimal projective
resolution. We remind the reader: a minimal projective resolution is a bounded
above chain complex of finitely generated, free R–modules, so that all the
differentials vanish modulo m. The object k L ⊗ Lx can be computed by taking
the ordinary tensor product of k with the minimal projective resolution. The
fact that k L ⊗ Lx ∼ = n k tells us that the minimal projective resolution must
be very sparse; it consists of a single free module, of rank 1, concentrated in
degree −n. That is Lx ∼ = n R.
So far we have focused on the stalk of a single point x ∈ X. We have a
chain complex L ∈ Db (Coh/X), that is a bounded chain complex of coherent
sheaves on X. We have shown that, at the point x, the stalks of the cohomology
sheaves Hk (L) vanish if k = −n, and if k = −n we obtained an isomorphism
of the stalk at x of H−n (L) with the free module R of rank 1. But the sheaves
318 Amnon Neeman
Hk (L) are coherent sheaves on X, all but finitely many of which vanish. It
follows that there exists a Zariski open set U ⊂ X so that, on U , the coherent
sheaves Hk (L) satisfy

0 if k = −n
Hk (L)|U =
L if k = −n
where L is some line bundle on U . Now the integer n is locally constant, and
must be constant on connected components. On each connected component Xi
there is an integer ni so that Hk (L) = 0 for k = −ni , and H−ni (L) is a line
bundle. This makes L|Xi a complex whose cohomology sheaves vanish except
in only one dimension, and hence it must be isomorphic in Db (Coh/Xi ) to
ni H−ni (L).
Remark 3.10. Let I be a dualizing complex. It is an object of Db (Coh/X), and
has a bounded-below injective resolution. In the homotopy category K(Inj/X),
whose objects are chain complexes of injective quasicoherent sheaves on X,
there is a bounded-below complex I and a quasi-isomorphism I −→ I . Fur-
thermore, I is unique up to homotopy. If J is another dualising complex,
Lemma 3.9 informs us that J = L L ⊗ I, with L locally isomorphic to a shift of
a line bundle. Obviously, L ⊗ I is an injective resolution for J.
While the injective resolutions I ∈ K(Inj/X) of dualizing complexes I are
only determined up to twisting by complexes L, the Hom–complexes Hom(I, I )
are objects of K(Flat/X) well defined up to homotopy; we have
Hom(I, I ) ∼
= Hom(L ⊗ I, L ⊗ I ).
Even though the dualizing complex I started its life as an object of the derived
category Db (Coh/X), subject to conditions that appear innocuous enough, we
have learned
(i) The object I is unique up to tensor by an L, with L locally isomorphic to
a shift of a line bundle.
(ii) If we replace I by its injective resolution, then the object Hom(I, I) is a
well-defined complex of flat OX –modules, unique up to homotopy.
In Proposition 3.6 we learned that the natural map OX −→ Hom(I, I) must be
an isomorphism in Db (Coh/X). It is a homology isomorphism of complexes
of flat modules, but not usually a homotopy equivalence.
Our next observation is
Theorem 3.11. Being a dualizing complex is local. In other words let Y = ∪Ui
be an open cover. Then a complex I ∈ D(Qcoh/Y ) is dualizing if and only if,
for each i, the restriction of I to Ui is dualizing.
Proof. First of all a point of notation: let fi : Ui −→ Y be the open immersion;

the functor fi∗ is exact, and hence is equal to its derived functor. We will write
fi∗ for Lfi∗ . Also, I belongs to Db (Coh/Y ) ⊂ D(Qcoh/Y ) if and only if each
fi∗ I belongs to Db (Coh/Ui ). We may therefore assume that I and all the fi∗ I
are bounded complexes of coherent sheaves.
By Proposition 3.6 it suffices to prove two things:
(i) The following are equivalent:

(a) For each object E ∈ Db (Coh/Y ) the complex RHom(E, I) belongs to
Db (Coh/Y ).
(b) For all i and all objects Ei ∈ Db (Coh/Ui ), the complexes
RHom(Ei , fi∗ I) belong to Db (Coh/Ui ).
(ii) The map OY −→ RHom(I, I) is an isomorphism if and only if, for each i,
the map OUi −→ RHom(fi∗ I, fi∗ I) is an isomorphism.
Let us begin with (i). The proof that (b)=⇒(a) is easy. Let E belong to
Db (Coh/Y ); by Lemma 2.4 we know that
RHom(fi∗ E, fi∗ I) = fi∗ RHom(E, I) .
By (b) the term on the left is in Db (Coh/Ui ) for every i, and from the iso-
morphism with the term on the right we conclude that RHom(E, I) must be in
Db (Coh/Y ).
Slightly subtler is (a)=⇒(b). Suppose (a) holds, and Ei is an object of
Db (Coh/Ui ). Since coherent sheaves and morphisms of coherent sheaves can
be extended from the open set Ui to all of Y , the object Ei ∈ Db (Coh/Ui ) is
isomorphic to fi∗ G, where G ∈ Db (Coh/Y ). Therefore
RHom(Ei , fi∗ I) = RHom(fi∗ G, fi∗ I)

= fi∗ RHom(G, I) by Lemma 2.4 .
Now RHom(G, I) is in Db (Coh/Y ) by (a), and f ∗ takes Db (Coh/Y ) to

Db (Coh/Ui ).
Next we prove (ii). Lemma 2.4 informs us that RHom(fi∗ I, fi∗ I) =
∗
fi RHom(I, I), and the map OY −→ RHom(I, I) will be an isomorphism if
and only if its restrictions to each Ui are.
Remark 3.12. Let f : X −→ Y be an open immersion. Then, for any dualizing

complex I on Y , the complex f ∗ I is dualizing on X. Just apply Theorem 3.11
to the open cover Y = X ∪ Y .
Remark 3.13. In Remark 3.12 we noted that dualizing complexes are preserved
when we pull them back by open immersions. It is natural to wonder what
happens with other functors.
320 Amnon Neeman
Let us make this precise. Suppose we are given a morphism of noethe-

rian, separated schemes f : X −→ Y ; there is a pair of derived categories
D(Qcoh/X) and D(Qcoh/Y ), and the morphism of schemes f : X −→ Y
induces several functors between these derived categories. There is the push-
forward functor Rf∗ : D(Qcoh/X) −→ D(Qcoh/Y ), which has a left adjoint
Lf ∗ and a right adjoint f ! . Assuming that either Db (Coh/X) or Db (Coh/Y )
has a dualizing complex, one can ponder whether some of the three functors
above might respect it. What we know so far, from Remark 3.12, is that, as
long as f is an open immersion, the functor Lf ∗ takes dualizing complexes on
Y to dualizing complexes on X. The next theorem gives sufficient conditions
for f ! to take a dualizing complex on Y to a dualizing complex on X.
Remark 3.14. Before we state the theorem, it might help if we remind the
reader of what boundedness properties are always, unconditionally true, for
any morphism f : X −→ Y ; this might help separate what is formal from the
genuinely restrictive hypotheses of the theorem.
We are assuming that X and Y are noetherian and separated. This means that
the functor Rf∗ can be computed using Čech cohomology with respect to an
affine open cover of X, and this cover may be taken finite. If E ∈ D(Qcoh/X)
is bounded it follows that Rf∗ E will also be a bounded complex. That much is
free. But we have no obvious way to control the size (as in coherence versus
quasicoherence) of the finitely many cohomology sheaves of Rf∗ E.
It turns out to be a strong restriction on f to assume that Rf∗ :
D(Qcoh/X) −→ D(Qcoh/Y ) respects the bounded, coherent subcategories.
In symbols: it may happen that Rf∗ will take the subcategory Db (Coh/X) ⊂
D(Qcoh/X) to the subcategory Db (Coh/Y ) ⊂ D(Qcoh/Y ), but it certainly does
not come for free. The usual sufficient condition is that the map f : X −→ Y
be a proper morphism of finite type. This concludes what we will need to know
concerning the functor Rf∗ , to make sense of the statement of Theorem 3.15;
the preservation of boundedness is formal, the preservation of coherence is not.
Next we want to look at the properties of the functor f ! which play a role
in the statement of Theorem 3.15. It helps to begin by returning briefly to
the functor Rf∗ , and observing a little more closely the precise bounds that
the Čech complex gives. Suppose that X admits a cover by n + 1 affine open
sets; then the Čech complex computing Rf∗ has length n. If E is an object in
D<m (Qcoh/X), that is E ∈ D(Qcoh/X) is quasi-isomorphic to a complex van-
ishing in degrees ≥ m, then Rf∗ E must be in D<m+n (Qcoh/Y ); a Čech complex
of length n can only raise the cohomological degree by at most n. This means
that for any F ∈ D≥m+n (Qcoh/Y ) we have Hom(Rf∗ E, F) = 0. Adjunction
gives us Hom(E, f ! F) = 0, and this is true for all E ∈ D<m (Qcoh/X). Hence
f ! F ∈ D≥m (Qcoh/X). We have just shown that the functor f ! must take
bounded-below complexes to bounded-below complexes. In particular the
image under f ! of the subcategory Db (Coh/Y ) ⊂ D(Qcoh/Y ) must lie in
D+ (Qcoh/X) ⊂ D(Qcoh/X).
That ends our free ride. There is no formal reason to expect the func-
tor f ! to take Db (Coh/Y ) ⊂ D(Qcoh/Y ) into Db (Coh/X), or even into the
larger D+ Coh (Qcoh/X). We remind the reader: the objects in the category
D+Coh (Qcoh/X) are the bounded-below chain complexes of quasicoherent
sheaves, whose cohomology is all coherent.
Next we prove:
Theorem 3.15. Let f : X −→ Y be a morphism of noetherian, separated

schemes so that
(i) The functor Rf∗ : D(Qcoh/X) −→ D(Qcoh/Y ) takes Db (Coh/X) ⊂

D(Qcoh/X) into Db (Coh/Y ) ⊂ D(Qcoh/Y ).
(ii) The functor f ! : D(Qcoh/Y ) −→ D(Qcoh/X) takes Db (Coh/Y ) ⊂
D(Qcoh/Y ) into D+
Coh (Qcoh/X) ⊂ D(Qcoh/X).
If I is a dualizing complex on Y then f ! I is a dualizing complex on X.
Remark 3.16. The reader might wish to compare Theorem 3.15 with older
results in the literature; see, for example, [15, Remark 2, p. 299], [66, Corol-
lary 3, p. 396], [2, Proposition 2.5.11] and [45, §9.1 and §9.2].
Reminder 3.17. In the proof of Theorem 3.15 we will appeal to the technical
result [46, Theorem 4.2]; let us therefore remind the reader. The technical result
asserts the following: in the category D(Qcoh/X) there is a compact generator
S which detects non-vanishing high cohomology. Precisely, this means that
there is an integer A = A(S), depending only on S, so that
(iii) If G is an object of D(Qcoh/X), with Hom(S, k G) = 0 for all k ≥ n,

then Hk (G) = 0 for all k ≥ n + A.
Perhaps we should explain the notation: the Hk means the kth cohomology
sheaf of the chain complex of sheaves. The result informs us that we can
tell whether the sheaf cohomology vanishes, above a certain degree, just by
computing some groups, namely Hom(S, k G).
Proof. We now prove Theorem 3.15. As in the proof of Theorem 3.11 we

appeal to Proposition 3.6. It suffices to establish two things:
322 Amnon Neeman
(iv) For all objects G ∈ Db (Coh/X) we will show that RHom(G, f ! I) belongs
to Db (Coh/X). The case where G = OX will prove that f ! I ∈ Db (Coh/X).
(v) We will show that the natural map OX −→ RHom(f ! I, f ! I) is an isomor-
phism.
Let us begin with (iv). What we are given is that G belongs to Db (Coh/X),
and hypothesis (ii) of the Theorem says that f ! I must belong to D+
Coh (Qcoh/X).
+
It follows that RHom(G, f I) has to be in DCoh (Qcoh/X). We remind the reader:
!
the complex G is finite, and hence we can immediately reduce to the case where
G is a single coherent sheaf concentrated in degree 0. If G is a single sheaf,
then there is a spectral sequence converging to the cohomology sheaves of the
ij ij
complex RHom(G, f I), whose E2 term is Ext G , H (f I) . The sheaves E2
! i j !
are all coherent, and Hi+j RHom(G, f ! I) is a finite extension of subquotients

of them, hence also coherent.
So much for coherence; to establish (iv) it remains to show that the cohomol-
ogy of RHom(G, f ! I) is bounded above. For this we use [46, Theorem 4.2].
Choose an object S as in Reminder 3.17. We know that the vanishing, for
sufficiently large k, of the two sequences

Hk RHom(G, f ! I) , Hom S , k RHom(G, f ! I)
is equivalent. And the point is that we can compute the groups on the right;
they are

Hom S , RHom(G , k f ! I) = Hom(S ⊗ GL, f I) k
L k !
= Hom Rf∗ (S ⊗ G), I .

Now S is a perfect complex; locally it is a bounded complex of finitely generated
projectives. The complex G is locally a bounded complex of finite modules, and
the derived tensor product is locally just the tensor product; locally S L ⊗ G is a
bounded complex of finite modules. Thus S L ⊗ G belongs to Db (Coh/X), and
by (i) it follows that Rf∗ (S L ⊗ G) must belong to Db (Coh/Y
I is
). The complex
by hypothesis a dualizing complex, and hence RHom Rf∗ (S ⊗ G), I must
L
also belong to Db (Coh/Y ). Its cohomology may be computed using the Čech
complex on Y , and hence is bounded. Therefore

Hom Rf∗ (S L ⊗ G), k I = H k RHom Rf∗ (S L ⊗ G), I
must vanish for k sufficiently large.

It remains to prove (v); we must show that the natural map ρ : OX −→
RHom(f ! I, f ! I) is an isomorphism. We have already proved, in (iv), that for
any G ∈ Db (Coh/X) we have RHom(G, f ! I) ∈ Db (Coh/X). If we put G = OX
we conclude that f ! I ∈ Db (Coh/X), and then if we put G = f ! I we deduce
that RHom(f ! I, f ! I) must also be in Db (Coh/X). The morphism ρ : OX −→

RHom(f ! I, f ! I) is therefore a map in Db (Coh/X), and the mapping cone lies
in Db (Coh/X). Locally the mapping cone has a minimal resolution; it will
vanish if and only if, for all closed points x ∈ X, the tensor product with k(x)
vanishes. To prove that ρ an isomorphism it therefore suffices to check that
k(x) L ⊗ ρ is an isomorphism for every closed point x ∈ X. That is, we wish to
study the natural map
k(x) −−−−→ k(x) L ⊗ RHom(f ! I, f ! I)
and prove it an isomorphism.
Lemma 2.10 helps. The point x is assumed closed, guaranteeing that the
skyscraper sheaf k(x), concentrated at the point x ∈ X, is a coherent sheaf.
Lemma 2.10 therefore gives us a natural isomorphism
. /
k(x) L ⊗ RHom(f ! I, f ! I) −−−−→ RHom RHom k(x) , f ! I , f ! I ,
and we are reduced to proving that the natural map

. /
k(x) −−−−→ RHom RHom k(x) , f ! I , f ! I ,
is an isomorphism. In the case of skyscraper sheaves concentrated at a single

point we can safely apply Rf∗ , and then check that the induced map on Y is
an isomorphism; after all the only effect is to push the problem forward to the
point f (x) ∈ Y . But now observe
. / . /
Rf∗ RHom RHom k(x) , f ! I , f ! I ∼ = RHom Rf∗ RHom k(x) , f ! I , I
. /
∼
= RHom RHom Rf∗ k(x) , I , I
∼
= Rf∗ k(x) .
The careful reader should note that, in the above isomorphisms, we assert that
two morphisms Rf∗ RHom(E, f ! I) −→ RHom(Rf∗ E, I) are isomorphisms. In
both cases E ∈ Db (Coh/X) is supported at a single point x ∈ X, meaning that
both complexes we are trying to prove isomorphic are acyclic away from
f (x) ∈ Y . The fact that the natural map is an isomorphism may therefore be
checked after taking global sections. The reader need not worry that we might
be appealing to some subtle facts about f ! commuting with base change, and
there is no need to verify the hypotheses of [51, Lemma 6.1 and Proposition 6.2].
The proof is now complete.
Remark 3.18. The hypothesis of Theorem 3.15 are that (i) Rf∗ take
Db (Coh/X) to Db (Coh/Y ), and (ii) f ! take Db (Coh/Y ) to D+Coh (Qcoh/X).
In Reminder 3.14 we noted that (i) holds provided f is a proper morphism of
324 Amnon Neeman
finite type. All we observed concerning (ii) was that it is not automatic; we
made no mention of any interesting examples of f s which satisfy (ii). It is time
to remedy this. We begin with the easy
Lemma 3.19. If f : X −→ Y is a finite morphism, then f ! : D(Qcoh/Y ) −→

D(Qcoh/X) takes Db (Coh/Y ) to D+
Coh (Qcoh/X).
Proof. If f is a finite morphism then it is affine; the sheaf f∗ OX is a coherent

sheaf of OY –algebras on Y , and X is simply Spec(f∗ OX ). The functor f∗ is
exact, and Rf∗ is just the forgetful functor, which takes a chain complex C of
OX –modules on X and views f∗ C as a chain complex of OY –modules via the
homomorphism OY −→ f∗ OX ; there is no need to derive the functor f∗ .
The right adjoint, or more precisely f∗ of the right adjoint, is the functor
which takes a complex D of OY modules on Y , and produces out of it the
complex RHomO (f∗ OX , D). Perhaps it might be clearer to restrict to an open
Y
affine of Y , giving the local description. Replacing Y by an open affine subset
and X by the inverse image, we have that Y = Spec(R) and X = Spec(S). The
morphism f : X −→ Y comes from a ring homomorphism ϕ : R −→ S, and
the fact that f is finite means that S is finite as an R–module. The functor f !
takes a complex D of R–modules to RHomR (S, D). We are assuming that D is a
bounded complex of finite R–modules, and wish to show that f ! D is a bounded-
below complex of S–modules, whose cohomology modules are finite over S.
Easy reduction tells us that we may assume the finite complex D is of length
1; we may take D to be a single, finite R–module concentrated in degree 0. The
cohomology of the complex f ! D is then Exti (S, D); it vanishes in negative
degrees, and is always finite, either as an R–module or as an S–module.
Example 3.20. Suppose Y is a any (noetherian, separated) scheme.

Then the sheaf OY certainly belongs to Db (Coh/Y ), and clearly satisfies
RHom(OY , OY ) = OY . This much is free.
If we assume that Y is finite dimensional and regular, or more generally that
it is Gorenstein, then the sheaf OY has finite injective dimension and therefore,
for every object E ∈ Db (Coh/Y ), we have that RHom(E, OY ) has only finitely
many non-vanishing cohomology sheaves. It is an easy exercise to show that
they are all coherent; therefore RHom(E, OY ) must belong to Db (Coh/Y ). We
conclude that OY is a dualizing complex for Y .
From Lemma 3.19 and Theorem 3.15 we know that, if f : X −→ Y is any
finite morphism, then f ! OY is a dualizing complex on X. The special case
where f : X −→ Y is a closed immersion tells us that any scheme, which
admits a closed immersion into a Gorenstein scheme, must have a dualizing
complex. Dualizing complexes are quite common.
4. When Rf∗ respects compacts

Let f : X −→ Y be a morphism of separated, noetherian schemes. Theo-
rem 3.15 gives us sufficient conditions for the functor f ! to take dualizing
complexes on Y to dualizing complexes on X. The conditions on f come
in two components: the first, part (i) of Theorem 3.15, says that the functor
Rf∗ should take Db (Coh/X) to Db (Coh/Y ). This hypothesis we understand;
see Remark 3.14. Part (ii) is the restriction that f ! should take Db (Coh/Y )
to D+ Coh (Qcoh/X). So far, the only example we have of an f satisfying (ii)
comes from Lemma 3.19; the hypothesis is satisfied for any finite morphism
f . It turns out that there is another large class of f s satisfying the condition
in Theorem 3.15(ii), and this section is devoted to studying them. We do not
understand this class fully; the unsatisfactory state of our current knowledge
will be made precise in Conjecture 4.16.
Consider the morphisms f : X −→ Y for which Rf∗ takes compacts to
compacts; in the literature they sometimes go by the name quasi-perfect. Sup-
pose Conjecture 4.16 is true. Assume f : X −→ Y is a quasi-perfect morphism
of noetherian, separated schemes, we will show that f ! takes Db (Coh/Y ) to
Db (Coh/X); it most certainly satisfies the hypothesis in Theorem 3.15(ii). In
other words, if the reader is willing to believe Conjecture 4.16, then Theo-
rem 3.15 applies to quasi-perfect f s.
There is a refinement; Conjecture 4.16 can be made separately for each
scheme Z. It suffices, in the paragraph above, to know that the conjecture is
true for X. Precisely: if f : X −→ Y is a quasi-perfect morphism of noetherian,
separated schemes, and if Conjecture 4.16 is true for X, then we already know
that f ! takes Db (Coh/Y ) to Db (Coh/X). In this section and in §5 we will explain
this, and show that there are many classes of Xs which satisfy Conjecture 4.16.
Quasi-perfect morphisms f have been studied extensively elsewhere, and
it seems remarkable that we will be able to say something new about them.
Because this section might be of independent interest, to people who could not
care less about dualizing complexes, we depart from our usual conventions in
the article. For this section we drop the hypothesis that our schemes should
be noetherian; during most of the section we will only assume them to be
quasicompact and separated.
Reminder 4.1. We remind the reader: if X is a quasicompact, separated scheme

then the category D(Qcoh/X) is compactly generated, and therefore has prod-
ucts; see [52, Proposition 8.4.6]. Next we note:
Lemma 4.2. Let f : X −→ Y be a morphism of quasicompact, separated

schemes, and assume Rf∗ : D(Qcoh/X) −→ D(Qcoh/Y ) takes compacts to
326 Amnon Neeman
compacts. Then the left adjoint Lf ∗ : D(Qcoh/Y ) −→ D(Qcoh/X) respects

products.
Proof. Let {Yλ , λ ∈ } be a set of objects in D(Qcoh/Y ). We wish to show
that the natural map
! "

∗
ϕ : Lf Yλ −−−−→ Lf ∗ Yλ
λ∈ λ∈
is an isomorphism. Since D(Qcoh/X) is compactly generated it suffices to

show that, for every compact object E ∈ D(Qcoh/X), the functor HomX (E, −)
takes ϕ to an isomorphism. Of course we can factor the functor HomX (E, −);
it can be expressed as a composite
RHom(E,−) Rf∗ H0
D(Qcoh/X) −−−−−−→ D(Qcoh/X) −−−−→ D(Qcoh/Y ) −−−−→ Ab,
and it clearly suffices to prove that the shorter composite
RHom(E,−) Rf∗
D(Qcoh/X) −−−−−−→ D(Qcoh/X) −−−−→ D(Qcoh/Y )
takes ϕ to an isomorphism. In Reminder 2.7 we saw that E is strongly dualizable
and hence the functor RHom(E, −) identifies with E∨ L ⊗ −. Parts (ii) and (iii)
of Reminder 2.7 assure us that E∨ is compact, while Reminder 2.7(i) informs
us that every compact object can be written as E∨ for some E. We are therefore
reduced to proving that, for any compact object E ∈ D(Qcoh/X), the composite
E L⊗ − Rf∗
D(Qcoh/X) −−−−→ D(Qcoh/X) −−−−→ D(Qcoh/Y )
takes ϕ to an isomorphism.
To do this, observe the isomorphisms
4 ! "5 ! "

Rf∗ E ⊗ Lf
L ∗
Yλ ∼ Rf∗ E ⊗
= L
Yλ projection formula
λ∈ λ∈

∼
= Rf∗ E L ⊗ Yλ Rf∗ E is compact
λ∈

∼
= Rf∗ E L ⊗ Lf ∗ Yλ projection formula
λ∈
L
∼
= Rf∗ E ⊗ Lf ∗ Yλ Rf∗ has left adjoint
4
λ∈ 5

∼ Rf∗ E L ⊗
= Lf ∗ Yλ E is compact .
λ∈
Note that in this string of isomorphisms we twice appealed to Reminder 2.7(iii);

tensor product with a compact object commutes with products. We used it for
the compact object E ∈ D(Qcoh/X) and for Rf∗ E ∈ D(Qcoh/Y ).
Remark 4.3. The fact that Lf ∗ respects products means that it must have a left
adjoint. Van den Bergh
* ! suggested+ a formula for this adjoint; it should be given
by the functor Rf∗ f OY ⊗ − . Lipman and Van den Bergh independently
L
found proofs that the formula works, at least for large classes of f s.
Lemma 4.4. Assume f : X −→ Y is a morphism of quasicompact, separated

schemes, and suppose further that Rf∗ takes compact objects in D(Qcoh/X)
to compact objects in D(Qcoh/Y ). Suppose {Xλ , λ ∈ } is a set of objects in
D(Qcoh/X), with each Xλ isomorphic to a coproduct of suspensions of OX .
Then the natural map
! "

f OY ⊗
! L
Xλ −−−−→ (f ! OY L ⊗ Xλ )
λ∈ λ∈
is an isomorphism.
Proof. We know that Lf ∗ n OY ∼ = n OX ; hence any suspension of OX can be

∗
expressed as Lf of some object of D(Qcoh/Y ), and so can any coproduct of
suspensions. For each of our objects Xλ ∈ D(Qcoh/X) we may choose an object
Yλ ∈ D(Qcoh/Y ) and an isomorphism Xλ ∼ = Lf ∗ Yλ . Now [51, Theorem 5.4]
gives us a natural isomorphism, for every Y ∈ D(Qcoh/Y ),
f !Y ∼
= f ! OY L ⊗ Lf ∗ Y ,
and we therefore have isomorphisms

! " ! "

f OY ⊗
! L ∼
Xλ = f OY ⊗
! L ∗
Lf Yλ
λ∈ λ∈! "

∼
= f O ⊗ Lf
! L ∗
Y Lemma 4.2
Y λ
! " λ∈

∼
= f! Yλ [51, Theorem 5.4]
λ∈

∼
= f ! Yλ f ! has left adjoint
λ∈

∼
= f ! OY L ⊗ Lf ∗ Yλ [51, Theorem 5.4]
λ∈

∼
= f ! OY L ⊗ Xλ ,
λ∈
completing the proof.

328 Amnon Neeman
Remark 4.5. We have shown that f ! OY is an object of D(Qcoh/X) so that the

functor f ! OY L ⊗ − commutes with some products; it commutes with products
of the Xλ s of Lemma 4.4. It is natural to ask whether it commutes with all
products. Reminder 2.7(iii) tells us that this is equivalent to asking whether
f ! OY is compact.
The general answer is No: we remind the reader how construct a coun-
terexample. Consider simple case where Y = Spec(k), where k is a field, and
f : X −→ Y is a finite morphism. For our choice of Y the sheaf OY is a
dualizing complex on Y and Lemma 3.19, coupled with Theorem 3.15, tell
us that f ! OY is a dualizing complex on X. Our hypothesis is that X is finite
over Y = Spec(k), in which case we know that the (essentially unique) dual-
izing complex has finite injective dimension. If it is also a compact object
then the complex RHom(f ! OY , f ! OY ) is quasi-isomorphic to the complex of
homomorphisms from a bounded complex of finitely generated projectives
to a bounded complex of injectives; but Proposition 3.6(ii) gives a quasi-
isomorphism OX −→ RHom(f ! OY , f ! OY ). This can only happen if OX has a
bounded injective resolution, that is if X is a Gorenstein scheme.
Remark 4.6. While f ! OY need not in general be compact, there are interesting
things one can say about it, facts which I do not fully understand. To illustrate
one of the strange features, a phenomenon which seems mysterious to me,
consider the following. By [51, Theorem 5.4] there is an isomorphism f ! (−) ∼ =
f ! OY L ⊗ Lf ∗ (−). Applying Rf∗ to this isomorphism, and using the projection
formula, we deduce an isomorphism
* +
Rf∗ f ! (−) ∼
= Rf∗ f ! OY L ⊗ − .
The functor on the left is the composite of two right adjoints, hence commutes
with products. It follows that so does the functor on the right; Reminder 2.7(iii)
now informs us that Rf∗ f ! OY has to be compact. Even though f ! OY need not
be compact, its pushforward Rf∗ f ! OY must be.
In the light of Lemma 4.4 and Remark 4.5 it makes sense to study the class
of objects L ∈ D(Qcoh/X) for which the functor L L ⊗ − commutes with the
products of Lemma 4.4. Let us make this a definition.
Definition 4.7. Let X be a quasicompact, separated scheme. We define a

class of objects S(X) ⊂ D(Qcoh/X) by the following property. An object L ∈
D(Qcoh/X) belongs to S(X) if and only if the natural map
! "

L ⊗
L
Xλ −−−−→ (L L ⊗ Xλ )
λ∈ λ∈
is an isomorphism whenever {Xλ , λ ∈ } is a set of objects in D(Qcoh/X) as

in Lemma 4.4. We remind the reader: this means that each Xλ isomorphic to a
coproduct of suspensions of OX .
Remark 4.8. When X can be understood from context we will omit it from
the notation; that is we will write S for S(X). Reminder 2.7(iii) tells us that
all compact objects belong to S = S(X). Lemma 4.4 says that, if f : X −→ Y
is a morphism where Rf∗ takes compacts to compacts, then f ! OY belongs
to S(X). Clearly, if E −→ F −→ G −→ E is a distinguished triangle, and
if two of the objects belong to S, then so does the third. Not so trivial is the
lemma
Lemma 4.9. Suppose E is a compact object in D(Qcoh/X), and let L be an

object in the class S of Definition 4.7. Then there exists a compact object F ∈
D(Qcoh/X), and a morphism f : F −→ L, so that any morphism n E −→ L
f
factors as n E −→ F −→ L.
Proof. Let be the set of all morphisms E −→ n L. If λ ∈ is a morphism

E −→ n L, put Xλ = n OX . Because L belongs to S we know that the natural
map
! "

L ⊗
L
Xλ −−−−→ (L L ⊗ Xλ )
λ∈ λ∈
is an isomorphism. The collection , of all maps E −→ n L = L L ⊗ Xλ ,

assembles to a single map

E −−−−→ (L L ⊗ Xλ ),
λ∈
which must therefore factor as

! "

E −−−−→ L L ⊗ Xλ −−−−→ (L L ⊗ Xλ ).
λ∈ λ∈
f L⊗ 1
But E is compact; any map E −→ G L ⊗ H factors as E −→ F L ⊗ H −→
G L ⊗ H, with F ∈ D(Qcoh/X) compact. Applying this to the above we deduce
a factorization
! " ! "
f L⊗ 1
E −−−−→ F ⊗ L
Xλ −−−−→ L ⊗ L
Xλ −−−−→ (L L ⊗ Xλ ).
λ∈ λ∈ λ∈
330 Amnon Neeman
Now consider the commutative diagram

! " ! "
f L⊗ 1
E −−−−→ F ⊗ L
Xλ −−−−→ L ⊗
L
Xλ
⏐
λ∈
⏐
λ∈
⏐ ⏐

(F L ⊗ Xλ ) −−−−−−→ (L L ⊗ Xλ ) .
λ∈ f ⊗ 1
L
λ∈ λ∈
By construction, the composite from top left to bottom right amounts to assem-
bling the collection of all maps E −→ n L into a single morphism. The
factorization
! "

E −−−−→ F L ⊗ Xλ
⏐λ∈
⏐

(F L ⊗ Xλ ) −−−−−−→ (L L ⊗ Xλ )
λ∈ f L⊗ 1
λ∈ λ∈
gives us a single map f : F −→ L, and informs us that each morphism E −→
n f
n L has a factorization E −→ n F −→ n L.
This already permits us to deduce the following two facts.

Lemma 4.10. Let X be a quasicompact, separated scheme. Let L be an object
belonging to S(X) ⊂ D(Qcoh/X). Then
(i) The object L is ℵ1 –compact, in the sense of [52, Definition 4.2.7].

(ii) If D(Qcoh/X) has a strong compact generator then L is compact.
Proof. By [5, Theorem 3.1.1] there is a compact generator in D(Qcoh/X); let

E be such a compact generator. If we are proving (ii), assume further that E is a
strong generator; this means that there exists an integer n so that any compact
can be obtained from E in n steps; see [5, Definition 2.2.3].
Lemma 4.4 tells us that we may choose a compact object F = F0 , and a
morphism f0 : F0 −→ L, so that any morphism n E −→ L factors through
f0 . That is, for every integer n the functor Hom( n E , −) takes f0 : F0 −→ L
to an epimorphism. Now assume we have a compact object Fi and a morphism
fi : Fi −→ L, and we will show how to produce a commutative triangle
Fi YYYYYY fi
YYYYYY
YYYY,
ϕi
eeeeee2 L
eeeee eeee
F e fi+1
i+1
with the properties that

(iii) The object Fi+1 is compact.
(iv) The group homomorphism Hom( n E , ϕi ) annihilates the kernel of
Hom( n E , fi ).
Complete fi to a distinguished triangle
g fi
L −−−−→ Fi −−−−→ L −−−−→ L .
By assumption L belongs to S(X), as does the compact object Fi . Therefore
L also belongs to S(X). Applying Lemma 4.4 to L we discover a compact
object G and a morphism h : G −→ L , so that every map n E −→ L factors
through h. That is, each of the functors Hom( n E , −) takes h to a surjection.
Since the functors Hom( n E , −) are homological they take any triangle to
exact sequences; for every integer n the functor Hom( n E , −) takes L −→
Fi −→ L to an exact sequence. Combining the surjections with the exact
sequences, we deduce the exactness of
Hom( n E , gh) Hom( n E , fi )
Hom( n E , G) −−−−−−−−→ Hom( n E , Fi ) −−−−−−−→ Hom( n E , L) .
The kernel of Hom( n E , fi ) is therefore the image of Hom( n E , gh). Now
complete gh : G −→ Fi to a distinguished triangle
gh ϕi
G −−−−→ Fi −−−−→ Fi+1 −−−−→ G .
Clearly Hom( n E , ϕi ) annihilates the image of Hom( n E , gh), which is the
kernel of Hom( n E , fi ); we have achieved (iv). Because G and Fi are both
gh
compact, so is Fi+1 ; that is (iii) also holds. Because the composite G −→
fi ϕi fi+1
Fi −→ L vanishes, the map fi must factor as Fi −→ Fi+1 −→ L. This yields
our commutative triangle
Fi YYYYYY fi
YYYYYY
YYYY,
ϕi
ee eeeeeee2 L
eeeee e
Fi+1 e fi+1
We have proved all our inductive claims.

We have a sequence of morphisms
ϕ0 ϕ1 ϕ2
F0 −−−−→ F1 −−−−→ F2 −−−−→ · · ·
and compatible maps to L; we can factor through a map ϕ : Hocolim - Fi −→

L. The usual argument, as in Brown’s original proof of the Brown representabil-
ity theorem, tells us that each of the functors Hom( n E , −) takes the map ϕ
332 Amnon Neeman
to an isomorphism; the reader can find this many places, for example in [51,
§3]. Since E generates it follows that ϕ is an isomorphism; L is isomorphic to
- fi , a countable homotopy colimit of compact objects. Hence L is
Hocolim
ℵ1 –compact. This establishes (i).
It remains to prove (ii); from now we assume E is a strong generator, and our
notation will be as in [5, §2.2]. Let n be an integer so that En = {D(Qcoh/X)}c .
Recall our sequence above, of morphisms fi : Fi −→ L. One can show, by an
induction on i which we leave to the reader,
(v) Any morphism K −→ L, with K ∈ Ei , can be factored through fi :

Fi −→ L.
fj
(vi) Given a vanishing composite K −→ Fj −→ L, with K ∈ Ei , then the
composite
ϕj ϕj +1 ϕj +2 ϕi+j +1
K −−−−→ Fj −−−−→ Fj +1 −−−−→ Fj +1 −−−−→ · · · −−−−→ Fi+j
already vanishes.
In our sequence we constructed a morphism f2n : F2n −→ L, with F2n com-

pact, that is with F2n ∈ En . By (v) it must factor through fn : Fn −→ L.
α fn ϕ
Choose a factorization F2n −→ Fn −→ L. The longer composite Fn −→
α fn
F2n −→ Fn −→ L is clearly equal to fn : Fn −→ L. Put e = 1 − αϕ. Then
e fn
e : Fn −→ Fn is a morphism so that the composite Fn −→ Fn −→ L vanishes.
e ϕ
By (vi) we deduce the vanishing of the composite Fn −→ Fn −→ F2n . This
e ϕ α
means that the longer composite Fn −→ Fn −→ F2n −→ Fn also vanishes;
that is e(1 − e) = 0. We have shown that e is an idempotent.
Therefore Fn splits as Fn ∼
= F ⊕ F , where αϕ = 1 on F and 0 on F . The
reader can now check that each of the functors Hom( n E , −) takes the map
F −→ L to an isomorphism, and we conclude that L ∼ = F is compact.
Remark 4.11. From Lemma 4.10(i) we learn that S(X) ⊂ {D(Qcoh/X)}ℵ1 ,

where {D(Qcoh/X)}ℵ1 is the full subcategory of ℵ1 –compact objects in
D(Qcoh/X). By [52, 8.4.2.2] we know that the category {D(Qcoh/X)}ℵ1 is
essentially small. We deduce that, up to isomorphisms, there is only a set of
objects in S(X).
Lemma 4.12. Let X = Spec(R) be an affine scheme. If L belongs to S(X) then

L must be isomorphic in D(Qcoh/X) ∼ = D(R–Mod) to a bounded above com-
plex of finitely generated projectives, and only finitely many of the cohomology
groups of L can be non-zero.
Proof. We apply Lemma 4.9 with E = R. We discover that there must exist
a single compact object F, that is a bounded complex of finitely generated,
projective R–modules, as well single map f : F −→ L, so that any morphism
n R −→ L factors through f . The morphisms n R −→ L are in bijection
with elements of H −n (L) and we learn that, for every i, the map H i (F) −→
H i (L) must be surjective. But F is bounded, hence H i (L) vanishes for all but
finitely many i. It remains to show that L is isomorphic in D(R–Mod) to a
bounded above chain complex of finitely generated projectives.
Suppose that i is the integer at which the cohomology of L stops; that
is H i (L) = 0, but H j (L) = 0 for all j > i. It suffices to produce a finitely
generated, projective module P i and an epimorphism P i −→ H i (L). Let us
first establish that this really suffices. Suppose such a morphism exists; it
must lift to a map in the derived category −i P i −→ L = L0 . The third
edge of the triangle −i P i −→ L −→ L1 −→ −i+1 P i is an object L1 ∈
S(X) whose cohomology stops at (i − 1). We can now apply induction to
obtain a sequence of finitely generated projectives and morphisms −k P k −→
Lk , where the cohomology of Lk stops at i − k, and these finitely generated
projectives assemble to a chain complex
· · · −→ P i−2 −→ P i−1 −→ P i −→ 0
quasi-isomorphic to L.
It therefore remains to prove the existence of the surjection P i −→ H i (L).
Let P be a K–projective resolution for L; we may choose P to be a bounded
above complex of projective modules, which stops at i. That is P is a chain
complex
· · · −→ Pi−2 −→ Pi−1 −→ Pi −→ 0
and H i (P) ∼
= H i (L) = 0. The morphism f : F −→ P is a map in the derived
category between bounded above chain complexes of projectives, and hence
may be realized by a chain map. That is, we have a chain map
∂ i+1
F
· · · −−−−−−→ F i−2 −−−−−−→ F i−1 −−−−−−→ F i −−−−−−→ F i+1 −−−−−−→ F i+2 −−−−−−→ · · ·
⏐ ⏐ ⏐ ⏐ ⏐
⏐ ⏐ ⏐ ⏐ ⏐

· · · −−−−−−→ Pi−2 −−−−−−→ Pi−1 −−−−−−→ Pi −−−−−−→ 0 −−−−−−→ 0 −−−−−−→ · · ·

∂i
P
From the first paragraph of the proof we know that the map H i (F) −→ H i (P)
is surjective; that is the kernel of ∂Fi+1 surjects onto the cokernel of ∂Pi . But then
Fi certainly must also surject to the cokernel; we have found our surjection
Fi −→ H i (P), with Fi finitely generated and projective.
334 Amnon Neeman
This ends what I can currently prove in glorious generality. Let me next
formulate a conjecture.
Conjecture 4.13. Let X be any quasicompact, separated scheme, and let L

be an object in S(X). Then L is pseudocoherent, and has only finitely many
non-vanishing cohomology groups.
Reminder 4.14. We remind the reader: a complex L is pseudocoherent pro-

vided the restriction of L, to any open affine subset U ⊂ X, is isomorphic in
D(Qcoh/U ) ∼ = D(R–Mod) to a bounded above complex of finitely generated
projective R–modules.
Remark 4.15. In view of Lemma 4.12, Conjecture 4.13 amounts to the state-
ment that if L ∈ S(X) and j : U −→ X is an open immersion of an affine open
subset, then j ∗ (L) is in S(U ).
Let us formulate an even stronger conjecture:
Conjecture 4.16. Let X be a quasicompact, separated scheme. Then X satisfies

the following:
(∗) For any quasicompact open subset U ⊂ X there exist a compact object
E ∈ D(Qcoh/X), and an integer n ≥ 1, so that
Rj∗ OU ∈ En ;
here j : U −→ X is the inclusion of the open set, and En is as in [5,

Definition 2.2.3].
Reminder 4.17. Perhaps we should remind the reader of the notation in [5,
Definition 2.2.3]. Let E be an object in some triangulated category T. The
full subcategory E1 is defined to be the one containing all direct summands
of all coproducts of arbitrary suspensions of E. The category En is defined
inductively; an object lies in En+1 if it is a direct summand of an object y,
and y fits in a triangle
x −−−−→ y −−−−→ z −−−−→ x
with x ∈ E1 and z ∈ En .
Remark 4.18. Now that we have reminded the reader of the notation in Con-
jecture 4.16, we should also explain its relevance. In the remainder of this
section we will show that Conjecture 4.16 implies Conjecture 4.13. In §5 we
will study the many cases in which we know Conjecture 4.16 to be true. And
then, in §6, we will return to the relation with Grothendieck duality.
Definition 4.19. Let X be a quasicompact, separated scheme. A class of objects

T ⊂ D(Qcoh/X) will be called adapted if the natural map
! "

L ⊗L
tλ −−−−→ (L L ⊗ tλ )
λ∈ λ∈
is an isomorphism, whenever the following conditions both hold:
(i) is a set of objects {tλ , λ ∈ }, where each tλ belongs to T .

(ii) The object L belongs to S(X).
Remark 4.20. By definition of S(X) the class T , containing all coproducts of

arbitrary suspensions of OX , is an adapted class. If T is an adapted class then
so is
(i) The class of all mapping cones on morphisms t −→ t , with t, t ∈ T .

(ii) The class of all direct summands of objects in T .
(iii) The class of all objects E L ⊗ t, where E is a fixed compact object and
t ∈ T.
All three facts are easy; we leave (i) and (ii) to the reader and indicate the proof
of (iii). Let E be a compact object, L an object in S, and assume all the tλ lie in
T . We have isomorphisms
4 5 ! "

L ⊗L
(E ⊗ t ) =
L ∼ L ⊗E ⊗
L L
t Reminder 2.7(iii)
λ λ
λ∈ 4 λ∈ 5

∼
= E L⊗ (L L ⊗ tλ ) Because L ∈ S, all tλ ∈ T
λ∈
∼
= (L L ⊗ E L ⊗ tλ ) Reminder 2.7(iii) .
λ∈
Lemma 4.21. Conjecture 4.16 implies Conjecture 4.13. More precisely for
each X, Conjecture 4.16 for X implies Conjecture 4.13 for X. We spell this
out: if X is a quasicompact, separated scheme, and if Conjecture 4.16 holds for
X, then every object L ∈ S(X) is pseudocoherent, and the cohomology sheaves
Hi (L) vanish for all but finitely many i.
Proof. Assume X is a quasicompact, separated scheme, let U ⊂ X be an affine

open set, and let j : U −→ X be the inclusion. We are assuming Conjec-
ture 4.16 is true; we may choose a compact object E ∈ D(Qcoh/X), and
an integer n ≥ 1, so that Rj∗ OU ∈ En . Choose and fix E and n. Note that
j : U −→ X is an affine morphism, hence j∗ is exact; there is no difference
between Rj∗ OU and j∗ OU . We will use the shorter j∗ OU .
336 Amnon Neeman
We observed, at the beginning of Remark 4.20, that the class T of all

coproducts of all suspensions of OX is adapted. By Remark 4.20(iii) so is
the class of all E L ⊗ t, with E as in the previous paragraph and t ∈ T . That
is, the class of all coproducts of arbitrary suspensions of E is adapted. By
Remark 4.20(ii) so is the class of all direct summands of the above; that is E1
is adapted. Now parts (i) and (ii) of Remark 4.20 tell us that if Ei is adapted
then so is E2i . We conclude that all the classes Ei are adapted.
The first paragraph of the proof informs us that En contains all coproducts
of arbitrary suspensions of j∗ OU . Now j∗ respects coproducts; a coproduct
of suspensions of j∗ OU is j∗ of the coproduct of suspensions of OU . Let
T ⊂ D(Qcoh/U ) be the class of all coproducts of arbitrary suspensions of
OU ; we have that j∗ T is an adapted class. This means that the natural map
! "

L L⊗ j∗ tλ −−−−→ (L L ⊗ j∗ tλ )
λ∈ λ∈
is an isomorphism, whenever the following conditions both hold:
(i) is a set of objects {tλ , λ ∈ }, where each tλ belongs to T .

(ii) The object L belongs to S(X).
Now take an object L ∈ S(X) and a set of objects {tλ , λ ∈ }, where each tλ
belongs to T , and observe the string of isomorphisms
4 ! "5 ! "

∗ L
j∗ j L ⊗ tλ ∼
= L ⊗ j∗
L
tλ projection formula
λ∈ ! λ∈ "

∼
= L L⊗ j∗ t j∗ has left adjoint
λ
λ∈
∼
= (L L ⊗ j∗ tλ ) j∗ T is adapted
λ∈

∼
= j∗ (j ∗ L L ⊗ tλ ) projection formula
λ∈4 5

∼
= j∗ ∗
(j L ⊗ tλ )
L
j∗ has left adjoint .
λ∈
Applying j ∗ to the isomorphism, and recalling that j ∗ j∗ is naturally isomorphic

to the identity, this simplifies to saying that the natural map
! "

∗ L
j L ⊗ tλ −−−−→ (j ∗ L L ⊗ tλ )
λ∈ λ∈
is an isomorphism. That is j ∗ L belongs to S(U ) ⊂ D(Qcoh/U ). Since U is

affine, Lemma 4.12 tells us that j ∗ L is a bounded above complex of finitely
generated projectives, with bounded cohomology.
5. Where we can prove Conjecture 4.16

This section is about studying the cases where we know Conjecture 4.16 to
be true. The conjecture is the assertion that sheaves of the form Rj∗ OU lie
in En , for suitable choices of a compact object E and an integer n. We
observe
Remark 5.1. Let X be a quasicompact, separated scheme. If there is a com-

pact object E ∈ D(Qcoh/X) and an integer n so that En = D(Qcoh/X), then
Conjecture 4.16 most definitely holds for X.
And the next observation is that there are examples in the literature. Specif-
ically, the reader is referred to the proof6 of [5, Theorem 3.1.4]. In there it is
shown that, if X is smooth over a field k, then En = D(Qcoh/X) for some n
and some compact E. In Theorem 5.13 we will prove a very slight refinement
of the result in [5]; there is no need to assume that the ground ring is a field, a
finite dimensional, regular, noetherian ring is quite enough.
This means that we have cheap examples of schemes X for which Conjec-
ture 4.16 is true; but all of them are smooth and noetherian. To obtain singular,
non-noetherian examples we will have to learn how to produce new examples
out of old ones. This section is mostly about developing the machinery.
We begin with a little definition.
Definition 5.2. Let X be a quasicompact, separated scheme. An open set

U ⊂ X will be called decent if
(i) U is quasicompact.
(ii) Let j : U −→ X be the inclusion. There exist a compact object E ∈
D(Qcoh/X), and an integer n ≥ 1, so that Rj∗ OU ∈ En .
Remark 5.3. In the terminology introduced in Definition 5.2, Conjecture 4.16

asserts that every quasicompact open subset U ⊂ X is decent. The idea of this
section will be to find ways to produce decent open subsets.
6 The reader should note that the statement of [5, Theorem 3.1.4] is slightly different; the result
we want is merely a step in the proof. See the last paragraph of [5, §3.4] for details.
338 Amnon Neeman
Lemma 5.4. Let X = Spec(R) be an affine scheme, and let Xf =

Spec(R[1/f ]) be the basic open subset consisting of the prime ideals not
containing f ∈ R. Then Xf ⊂ X is decent.
Proof. Being affine, Xf is clearly quasicompact. In the category D(Qcoh/X) ∼
=
D(R–Mod), the object Rj∗ OXf simplifies to R[1/f ] ∈ D(R–Mod). We need
find a compact object E and an integer n, and exhibit R[1/f ] as an object in
En . Now note that R[1/f ] is the colimit of the sequence
f f f f f
R −−−−→ R −−−−→ R −−−−→ R −−−−→ R −−−−→ · · ·
It is therefore also the homotopy colimit. There is a triangle in D(R–Mod)
∞
& ∞
& ∞
&
R −−−−→ R −−−−→ R[1/f ] −−−−→ R.
n=0 n=0 n=0
Put E = R and n = 2; the triangle shows that R[1/f ] belongs to R2 .

Lemma 5.5. Suppose X is a quasicompact, separated scheme. Let U and V
be open subsets, and assume that U , V and U ∩ V are all decent. Then so is
U ∪V.
Proof. The quasicompactness is clear; U ∪ V is the union of two quasicompact
open sets U and V . We have to worry about the sheaf Rj∗ OU ∪V . Since there
are several open subsets W ⊂ X to consider, and several inclusions among
them, we return to our notation of the proof of Lemma 2.4; given open subsets
W1 ⊂ W2 of X, we will denote the inclusion by jWW12 : W1 −→ W2 .
By hypothesis we may choose an integer n, and compact objects EU , EV
and EU ∩V in D(Qcoh/X), so that
R{jUX }∗ OU ∈ EU n , R{jVX }∗ OV ∈ EV n , R{jUX∩V }∗ OU ∩V ∈ EU ∩V n .
Put E = EU ⊕ EV ⊕ EU ∩V . Then E is compact, and the three complexes
R{jUX }∗ OU , R{jVX }∗ OV , R{jUX∩V }∗ OU ∩V
all belong to En . Now consider the distinguished triangle
R{jUX∪V }∗ OU ∪V −→ R{jUX }∗ OU ⊕ R{jVX }∗ OV −→
R{jUX∩V }∗ OU ∩V −→
Two of the objects belong to En and hence the third, namely R{jUX∪V }∗ OU ∪V ,
must belong to E2n .
Lemma 5.6. Let X = Spec(R) be affine. Then any quasicompact open subset
U ⊂ X is decent. That is, Conjecture 4.16 holds for X.
Proof. The open sets Xf , f ∈ R form a basis for the topology of X = Spec(R);
any open set U is the union of the Xf ’s contained in it. If U is quasicompact
we may express it as a finite union of Xf ’s. We will prove, by induction on n,
(i) The union of n basic open sets, that is open sets of the form Xf , is a decent
open set.
Lemma 5.4 gives us the case n = 1. Suppose therefore that we know (i) for n,
that U = ∪ni=1 Xfi is the union of n open sets Xfi , and that V = Xg is another
basic open set. We want to show that U ∪ V is decent, and Lemma 5.5 tells us
that it suffices to show that U , V and U ∩ V are. The case of V = Xg comes
from Lemma 5.4, while for U and
! n "
n n
U ∩V = Xfi ∩ Xg = (Xfi ∩ Xg ) = Xfi g
i=1 i=1 i=1
we appeal to the fact that both can be covered by n basic open sets.
Lemma 5.7. Decency is transitive. Precisely: suppose X is a quasicompact,

separated scheme, U ⊂ X is a decent open set of X, and V ⊂ U is a decent
open subset of U . Then V is a decent open subset of X.
Proof. The quasicompactness is clear. To keep the notation uncluttered let us

write j1 : V −→ U , j2 : U −→ X and j : V −→ X for the three inclusions;
we wish to prove that Rj∗ OV is contained in some K .
Because V is a decent open subset of U we may produce a compact object
E ∈ D(Qcoh/U ), and an integer m ≥ 1, so that R{j1 }∗ OV ∈ Em . The com-
plex E ⊕ E vanishes in the Grothendieck group K0 (U ), and Thomason’s
localization theorem [62, 5.2.2(a)] applies. There exists a compact object
F ∈ D(Qcoh/X) with j2∗ F ∼ = E ⊕ E. Choose such an F; we have that
R{j1 }∗ OV ∈ j2∗ Fm . (†)
We also know that U is decent in X. We may therefore find a compact object

G ∈ D(Qcoh/X), and an integer n ≥ 1, so that R{j2 }∗ OU ∈ Gn . Now we
combine these. If we apply the functor R{j2 }∗ to the inclusion (†) above, we
conclude that the complex Rj∗ OV = R{j2 }∗ R{j1 }∗ OV is contained in
R{j2 }∗ j2∗ Fm ⊂ R{j2 }∗ j2∗ Fm R{j2 }∗ respects coproducts

= R{j2 }∗ (j2∗ F L ⊗ OU )m obvious
= F L ⊗ R{j2 }∗ OU m projection formula
⊂ F L ⊗ Gmn ,
340 Amnon Neeman
where the last inclusion is because R{j2 }∗ OU belongs to Gn , hence

F L ⊗ R{j2 }∗ OU belongs to F L ⊗ Gn , and therefore
F L ⊗ R{j2 }∗ OU m ⊂ F L ⊗ Gmn .
Lemma 5.8. Let X be a quasicompact, separated scheme, and let {Ui , 1 ≤

i ≤ n} be a finite number of decent open affine subsets of X. Then the union
∪ni=1 Ui is decent.
Proof. We prove this by induction on n, the case n = 1 being trivial. Suppose

we know the assertion for n, and suppose we have n + 1 decent open affines
{Ui , 1 ≤ i ≤ n + 1}. Let U = ∪ni=1 Ui and let V = Un+1 ; by induction we know
that U and V are decent, and we want to prove that so is U ∪ V . By Lemma 5.5 it
suffices to prove the decency of U ∩ V . Now X is assumed separated, hence the
intersection of two quasicompact open subsets is quasicompact. Therefore U ∩
V ⊂ V is a quasicompact open subset of the affine scheme V , and Lemma 5.6
informs us that U ∩ V is decent in V . We are given that V is decent in X and the
transitivity of decency, that is Lemma 5.7, permits us to conclude that U ∩ V
is decent in X.
Proposition 5.9. Suppose X is a quasicompact, separated scheme, and suppose

that X can be covered by decent open affines. Then Conjecture 4.16 holds for
X; all quasicompact open sets are decent.
Proof. We are assuming that X is quasicompact and has a cover by decent open
affines. There must exist a finite subcover. We may cover X by n decent open
affines, for some integer n ≥ 1; the proof will be by induction on n. The case
n = 1 comes from Lemma 5.6. Suppose therefore that we know the assertion
for schemes X which admit covers by n decent open affines, and let X be
a scheme admitting a cover by n + 1 decent open affines. Then X = U ∪ V ,
where U has a cover by n decent open affines and V is a decent open affine.
Lemma 5.8 informs us that U ⊂ X is decent. Let W ⊂ X be any quasicompact
open subset; we want to prove that W is decent.
To do this observe that W can be written as the union
W = (W ∩ U ) ∪ (W ∩ V ) .
Now X is assumed separated, and hence the intersection of any two quasicom-
pact open sets is quasicompact; we deduce that W ∩ U , W ∩ V and W ∩ U ∩ V
are all quasicompact. Induction tells us that
(i) W ∩ U is decent as an open subset of the scheme U .

(ii) W ∩ V and W ∩ U ∩ V are both decent as open subsets of the scheme V .
Now the transitivity of decency, that is Lemma 5.7, guarantees that W ∩ U ,

W ∩ V and W ∩ U ∩ V are all decent as open subsets of X. Lemma 5.5 permits
us to conclude that W ⊂ X is decent.
Lemma 5.10. Let f : X −→ Y be a morphism of quasicompact, separated

schemes. If V ⊂ Y is decent then so is f −1 V ⊂ X.
Proof. Y is separated and the schemes X and V ⊂ Y are quasicompact, hence

f −1 V ⊂ X is quasicompact. The decency of V ⊂ Y says further that we
may choose a compact object E ∈ D(Qcoh/Y ) and an integer n ≥ 1 with
R{jVY }∗ OV ∈ En . Therefore Lf ∗ E is a compact object in D(Qcoh/X), and
Lf ∗ En contains the object
Lf ∗ R{jVY }∗ OV ∼
= R{jfX−1 V }∗ Lf ∗ OV ∼
= R{jfX−1 V }∗ Of −1 V .
Corollary 5.11. Let f : X −→ Y be a morphism of quasicompact, separated

schemes. Suppose X can be covered by open affine subsets f −1 V , where V ⊂ Y
is open and quasicompact. If Conjecture 4.16 holds for Y then it also holds
for X.
Proof. Conjecture 4.16 holds for Y and therefore any quasicompact V ⊂ Y

is decent. By Lemma 5.10 all the open sets f −1 V ⊂ X are decent, and by
hypothesis we may choose among them a collection of affines which cover X.
Proposition 5.9 allows us to conclude that Conjecture 4.16 holds for X.
Remark 5.12. There are two useful situations where Corollary 5.11 applies.
They are:
(i) Any open immersion of quasicompact, separated schemes.

(ii) Affine morphisms; for example closed immersions, or more generally finite
maps.
So much for machinery of confirming the decency of open subsets. Now we

come to the way of producing examples.
Theorem 5.13. Let R be a (noetherian) regular ring of finite global dimen-

sion, and assume X is a noetherian, separated scheme, smooth and finite
dimensional over R. There exists an integer n ≥ 1, as well as a compact object
E ∈ D(Qcoh/X), with En = D(Qcoh/X).
342 Amnon Neeman
Proof. As we already mentioned, the proof is a miniscule modification of an

argument that may be found in [5]. Consider the product X ×R X. It is a
finite dimensional, regular, noetherian scheme, and the sheaf O , that is the
structure sheaf of the diagonal embedding X −→ X ×R X, is a coherent sheaf
on X ×R X. Locally it has a finite resolution by finitely generated projectives;
that is O is a compact object in D(Qcoh/X ×R X).
Now choose a compact object E generating D(Qcoh/X); such an object
exists by [5, Theorem 3.1.1(2)]. From [5, Lemma 3.4.1] we learn that F =
π1∗ E L ⊗ π2∗ E is a compact generator of D(Qcoh/X ×R X). The compact object
O therefore lies in Fn for some integer n ≥ 1. Also, because R is a regular,
finite dimensional ring, there is an integer m with Rm = D(R–Mod). We
assert that Emn = D(Qcoh/X).
To prove this, choose any object S ∈ D(Qcoh/X). Then S can be expressed
as
* +
S∼= {π1 }∗ O L ⊗ π2∗ S .
We know that O ∈ Fn , with F = π1∗ E L ⊗ π2∗ E. Hence S must belong to

Gn , where
* + * +
G∼ = {π1 }∗ F L ⊗ π2∗ S ∼ = {π1 }∗ π1∗ E L ⊗ π2∗ E L ⊗ π2∗ S .
Manipulating a little more we have

* + * +
G∼ = {π1 }∗ π1∗ E L ⊗ π2∗ E L ⊗ S ∼ = E L ⊗ {π1 }∗ π2∗ E L ⊗ S ,
* +
and {π1 }∗ π2∗ E L ⊗ S ∼ = π ∗ π∗ E L ⊗ S is π ∗ of a complex π∗ E L ⊗ S
belonging to D(R–Mod) = Rm . Thus π ∗ π∗ E L ⊗ S must belong to OX m ⊂
D(Qcoh/X), and
* +
G∼= E L ⊗ π ∗ π∗ E L ⊗ S
lies in Em . Hence S ∈ Gn ⊂ Emn .
6. Dualizing complexes and f !

Let us leave the world of gorgeous generality and return to dealing with
more restricted schemes, at least some of which will be noetherian. Assume
f : X −→ Y is a morphism of quasicompact, separated schemes, suppose
Y is noetherian, and let I be a dualizing complex on Y . If E is any
object in D(Qcoh/X), and F belongs to Db (Coh/Y ) ⊂ D(Qcoh/Y ), then we
compute
( )
HomY (Rf∗ E, F) ∼
= HomY Rf∗ E , RHomY RHomY (F, I) , I
* +
∼
= HomY Rf∗ E L ⊗ RHomY (F, I), I
( )
∼
= HomY Rf∗ E L ⊗ Lf ∗ RHomY (F, I) , I
* +
∼
= HomX E L ⊗ Lf ∗ RHomY (F, I), f ! I
( )
∼
= HomX E , RHomX Lf ∗ RHomY (F, I) , f ! I ;
the first isomomorphism is by Lemma 3.5(ii), the second because RHom is right
adjoint to the tensor, the third by the projection formula, the fourth because
f ! is right adjoint to Rf∗ , and the fifth by the adjunction between tensor and
RHom. Put together, we deduce a natural isomorphism

f !F ∼
= RHomX Lf ∗ RHomY (F, I) , f ! I . (††)
We have found a formula (††) for computing f ! F, as long as F belongs to

Db (Coh/Y ). In the old days the right hand side was used to construct the
functor f ! ; in our modern day and age this seems anachronistic. We know,
for purely formal reasons, that the functor f ! : D(Qcoh/Y ) −→ D(Qcoh/X)
exists; see [51, Example 4.2].
Next we restrict our attention to an even smaller class of morphisms f :
X −→ Y ; for the rest of the section assume f is a morphism of noetherian,
separated schemes, and suppose further that
(i) Rf∗ takes compacts to compacts.

(ii) The scheme X satisfies Conjecture 4.16; for all we know this is no restric-
tion.
From [46, Corollary 4.3.2] it follows that
(iii) Rf∗ takes Db (Coh/X) to Db (Coh/Y ).
Lemma 4.4 informs us that f ! OY belongs to S(X), and, using (i) and (ii) above
as well as Lemma 4.21, we conclude that f ! OY is in D− (Coh/X). We therefore
deduce
(iv) f ! takes Db (Coh/Y ) to Db (Coh/X); see Remark 0.6 and Corollary 0.7.
(v) If I is a dualizing complex on Y , then f ! I is a dualizing complex on X;
see Fact 0.3(ii).
If we look at (iii) and (iv), they tell us that the adjoint pair of functors (Rf∗ , f ! )
restrict to functors between the subcategories Db (Coh/X) ⊂ D(Qcoh/X) and
344 Amnon Neeman
Db (Coh/Y ) ⊂ D(Qcoh/Y ); the restrictions must also be an adjoint pair. The

formula (††) even tells us how to compute f ! in terms of dualizing complexes.
In practice this means that, if it happens to be more convenient to compute
f ! in the bounded derived categories, then go right ahead; from the restriction
of f ! to Db (Coh/Y ) we can compute f ! OY , and the formula
f ! S = f ! OY L ⊗ Lf ∗ S
tells us, at least in principle, how to work out all there is to know about the
functor f ! : D(Qcoh/Y ) −→ D(Qcoh/X).
7. Several recent results

Let me end with a very brief glimpse at current progress in the field. We begin
by observing that, up to now, our treatment has been based on studying the
category Db (Coh/X) via its embedding in D(Qcoh/X). Dualizing complexes
naturally live in Db (Coh/X), and our study of them was by means of infinite
coproducts, infinite products and compact objects in D(Qcoh/X). As a result of
ongoing work, we now know that there are more natural compactly generated
triangulated categories to consider.
Krause [35] taught us that Db (Coh/X) can be viewed as the subcategory of
compact objects in the compactly generated triangulated category K(Inj/X).
From the work of Jørgensen [29], Iyengar-Krause [27], myself [53, 54] and
from Murfet’s thesis [48], we also know a compactly generated triangulated
category T = Km (Proj/X) whose subcategory T c of compact objects is natu-
rally equivalent to Db (Coh/X) . If a dualizing complex I exists, then tensor
op
product with (an injective resolution of) I gives an equivalence of categories
− ⊗ I : Km (Proj/X) −−−−→ K(Inj/X).
This approach is very new, poorly understood, and in my opinion it has great
promise. Limitations of space prevent me from discussing it any further here.
Appendix A. A fact concerning strongly dualizable objects

In this appendix we prove the assertion of Reminder 2.7(iii); we recall the
statement.
Theorem A.1. Let T be a compactly generated triangulated category, pos-

sessing a symmetric tensor product compatible with the triangulated structure.
Assume that the unit of the tensor is compact, and that all compact objects are
strongly dualizable. If E ∈ T is some object, then the following are equivalent:
(i) E is compact.
(ii) E is strongly dualizable.
(iii) Tensor product with E commutes with arbitrary products in T; that is, the
natural map

E∧ tλ −−−−→ (E ∧ tλ )
λ∈ λ∈
is an isomorphism for every set of objects {tλ , λ ∈ }.

Remark A.2. The implication (i)=⇒(ii) follows immediately from the hypoth-
esis of the theorem. The implications (ii)=⇒(i) and (ii)=⇒(iii) are known; for
(ii)=⇒(i) see [19, Theorem 2.1.3], while (ii)=⇒(iii) may be found in [19,
Theorem A.2.5(f)]. The fact that seems new is the implication (iii)=⇒(i). For
the convenience of the reader we give a complete, self-contained proof of the
equivalence of the three conditions.
Also, because this appendix might be of interest to people who could not
care less about the categories D(Qcoh/X), our notation will be the standard
one in the literature. The tensor product of two objects E, G ∈ T will be written
E ∧ G, the unit of the tensor will be denoted S, and the internal Hom-object
will be F (E, G). To translate back to the case where T = D(Qcoh/X), put
S = OX , E ∧ G = E L⊗ G , F (E, G) = RHom(E, G) .
If you read again the statement of Theorem A.1 you will observe that we already
used this notation in part (iii) of the theorem. With the notation established, it
is time to come to the proof of Theorem A.1.
Proof. We are assuming that the compact objects are all strongly dualizable;
thus (i)=⇒(ii) is trivial. Next we prove (ii)=⇒(iii). Suppose therefore that E is
strongly dualizable. In Reminder 2.7(i) we noted that there is an isomorphism
E∼ ∨
= {E∨ } , and in Reminder 2.7(ii) we observed that E∨ is strongly dualiz-
able. Putting this together we have that E ∧ G ∼ ∨
= {E∨ } ∧ G must be naturally
isomorphic to F (E∨ , G). We are therefore reduced to proving that the func-
tor F (E∨ , −) respects products. But this is obvious; F (E∨ , −) is right adjoint
to − ∧ E∨ .
It remains to prove (iii)=⇒(i), which is the part that seems new. Consider the
functor Hom(S , E ∧ −). The functor E ∧ − respects products by hypothesis
and coproducts obviously. The functor Hom(S, −) respects products obviously
and coproducts because we are assuming S compact. Therefore the composite
functor Hom(S , E ∧ −) must respect both products and coproducts.
346 Amnon Neeman
The fact that it respects products means that it must be representable; see [52,
Theorem 8.6.1]. There is an object G ∈ D(Qcoh/X) and a natural isomorphism
ϕ : Hom(G, −) −−−−→ Hom(S , E ∧ −) .

Because the functor Hom(S , E ∧ −) respects coproducts so does the isomor-
phic functor Hom(G, −), meaning that G must be compact, and therefore also
strongly dualizable. Yoneda’s lemma tells us that the isomorphism ϕ must come
from an element in

Hom(S , E ∧ G) ∼= Hom S , F (G∨ , E) ∼ = Hom(G∨ , E) .
This produces for us a morphism α : G∨ −→ E, which we will prove an iso-
morphism. What we know is that the natural map
Hom(S , α ∧ −) : Hom(S , G∨ ∧ −) −−−−→ Hom(S , E ∧ −)

is an isomorphism; it is just the map ϕ. In particular, for every compact object
K we deduce that
Hom(S , α ∧ 1) : Hom(S , G∨ ∧ K∨ ) −−−−→ Hom(S , E ∧ K∨ )

is an isomorphism, but this identifies with
Hom(K , α) : Hom(K , G∨ ) −−−−→ Hom(K , E) .

The compacts K generate, hence α must be an isomorphism, making E iso-
morphic to the object G∨ . We know that G is compact; it remains to prove that
so is G∨ . But we have a natural isomorphism
Hom(S , G ∧ −) ∼
= Hom(G∨ , −) ,
and the functor on the left clearly respects coproducts.
References
[1] Leovigildo Alonso Tarrı́o, Ana Jeremı́as López, and Joseph Lipman, Local homol-
ogy and cohomology on schemes, Ann. Sci. École Norm. Sup. (4) 30 (1997), no. 1,
1–39.
[2] , Studies in duality on Noetherian formal schemes and non-Noetherian
ordinary schemes, American Mathematical Society, Providence, RI, 1999.
[3] Leovigildo Alonso Tarrı́o, Ana Jeremı́as López, and Marı́a José Souto Salorio,
Localization in categories of complexes and unbounded resolutions, Canad. J.
Math. 52 (2000), no. 2, 225–247.
[4] Marcel Bökstedt and Amnon Neeman, Homotopy limits in triangulated categories,
Compositio Math. 86 (1993), 209–234.
[5] Alexei I. Bondal and Michel van den Bergh, Generators and representability of
functors in commutative and noncommutative geometry, Mosc. Math. J. 3 (2003),
no. 1, 1–36, 258.
[6] Brian Conrad, Grothendieck duality and base change, Lecture Notes in Mathe-
matics, vol. 1750, Springer-Verlag, Berlin, 2000.
[7] Pierre Deligne, Cohomology à support propre en construction du foncteur f ! ,
Residues and Duality, Lecture Notes in Mathematics, vol. 20, Springer–Verlag,
1966, pp. 404–421.
[8] , Cohomologie à supports propres, Théorie des topos et cohomologie étale
des schémas. Tome 3 (Berlin), Lecture Notes in Mathematics, vol. 305, Springer-
Verlag, 1973, (SGA 4, Exposé XVII), pp. 250–461.
[9] Duiliu-Emanuel Diaconescu, Enhanced D-brane categories from string field the-
ory, J. High Energy Phys. (2001), no. 6, Paper 16, 19.
[10] Albrecht Dold and Dieter Puppe, Duality, trace, and transfer, Proceedings of
the International Conference on Geometric Topology (Warsaw, 1978) (Warsaw),
PWN, 1980, pp. 81–102.
[11] Fouad El Zein, Complexe dualisant et applications à la classe fondamentale d’un
cycle, Bull. Soc. Math. France Mém. (1978), no. 58, 93.
[12] Pierre-Paul Grivel, Une démonstration du théorème de dualité de Verdier, Enseign.
Math. (2) 31 (1985), no. 3-4, 227–247.
[13] Alexandre Grothendieck, The cohomology theory of abstract algebraic varieties,
Proc. Internat. Congress Math. (Edinburgh, 1958), Cambridge Univ. Press, New
York, 1960, pp. 103–118.
[14] , Théorèmes de dualité pour les faisceaux algébriques cohérents,
Séminaire Bourbaki, Vol. 4, Soc. Math. France, Paris, 1995, pp. Exp. No. 149,
169–193.
[15] Robin Hartshorne, Residues and duality, Lecture Notes in Mathematics, vol. 20,
Springer–Verlag, 1966.
[16] Glenn W. Hopkins, An algebraic approach to Grothendieck’s residue symbol,
Trans. Amer. Math. Soc. 275 (1983), no. 2, 511–537.
[17] Glenn W. Hopkins and Joseph Lipman, An elementary theory of Grothendieck’s
residue symbol, C. R. Math. Rep. Acad. Sci. Canada 1 (1978/79), no. 3, 169–
172.
[18] Kentaro Hori and Johannes Walcher, D-branes from matrix factorizations, C. R.
Phys. 5 (2004), no. 9-10, 1061–1070, Strings 04. Part I.
[19] Mark Hovey, John H. Palmieri, and Neil P. Strickland, Axiomatic stable homotopy
theory, vol. 128, Memoirs AMS, no. 610, Amer. Math. Soc., 1997.
[20] Reinhold Hübl and Ernst Kunz, Integration of differential forms on schemes, J.
Reine Angew. Math. 410 (1990), 53–83.
[21] , Regular differential forms and duality for projective morphisms, J. Reine
Angew. Math. 410 (1990), 84–108.
[22] Reinhold Hübl and Pramathanath Sastry, Regular differential forms and relative
duality, Amer. J. Math. 115 (1993), no. 4, 749–787.
[23] Luc Illusie, Conditions de finitude, Théorie des intersections et théorème
de Riemann-Roch, Springer-Verlag, Berlin, 1971, Séminaire de Géométrie
Algébrique du Bois-Marie 1966–1967 (SGA 6, Exposé III), pp. 222–273. Lecture
Notes in Mathematics, Vol. 225.
348 Amnon Neeman
[24] , Existence de résolutions globales, Théorie des intersections et théorème

de Riemann-Roch, Springer-Verlag, Berlin, 1971, Séminaire de Géométrie
Algébrique du Bois-Marie 1966–1967 (SGA 6, Exposé II), pp. 160–221. Lec-
ture Notes in Mathematics, Vol. 225.
[25] , Généralités sur les conditions de finitude dans les catégories dérivées,
Théorie des intersections et théorème de Riemann-Roch, Springer-Verlag, Berlin,
1971, Séminaire de Géométrie Algébrique du Bois-Marie 1966–1967 (SGA 6,
Exposé I), pp. 78–159. Lecture Notes in Mathematics, Vol. 225.
[26] , Groupes de grothendieck des topos annelés, Théorie des intersections
et théorème de Riemann-Roch, Springer-Verlag, Berlin, 1971, Séminaire de
Géométrie Algébrique du Bois-Marie 1966–1967 (SGA 6, Exposé IV), pp. 274–
296. Lecture Notes in Mathematics, Vol. 225.
[27] Srikanth Iyengar and Henning Krause, Acyclicity versus total acyclicity for com-
plexes over Noetherian rings, Documenta Math. 11 (2006), 207–240.
[28] Peter Jørgensen, Serre-duality for tails(A), Proc. Amer. Math. Soc. 125 (1997),
no. 3, 709–716.
[29] , The homotopy category of complexes of projective modules, Adv. Math.
193 (2005), no. 1, 223–232.
[30] Anton N. Kapustin and Yi Li, D-branes in Landau-Ginzburg models and algebraic
geometry, J. High Energy Phys. (2003), no. 12, 005, 44 pp. (electronic).
[31] Anton N. Kapustin and Dmitri O. Orlov, Lectures on mirror symmetry, derived
categories, and D-branes, Uspekhi Mat. Nauk 59 (2004), no. 5(359), 101–134.
[32] Bernhard Keller, Deriving DG categories, Ann. Sci. École Norm. Sup. (4) 27
(1994), no. 1, 63–102.
[33] G. Maxwell Kelly, Many-variable functorial calculus. I, Coherence in categories,
Springer, Berlin, 1972, pp. 66–105. Lecture Notes in Math., Vol. 281.
[34] Reinhardt Kiehl, Ein “Descente”-Lemma und Grothendiecks Projektionssatz für
nichtnoethersche Schemata, Math. Ann. 198 (1972), 287–316.
[35] Henning Krause, The stable derived category of a Noetherian scheme, Compos.
Math. 141 (2005), no. 5, 1128–1162.
[36] Ernst Kunz, Residuen von Differentialformen auf Cohen-Macaulay-Varietäten,
Math. Z. 152 (1977), no. 2, 165–189.
[37] , Über den n-dimensionalen Residuensatz, Jahresber. Deutsch. Math.-
Verein. 94 (1992), no. 4, 170–188.
[38] , Geometric applications of the residue theorem on algebraic curves,
Algebra, arithmetic and geometry with applications (West Lafayette, IN, 2000),
Springer, Berlin, 2004, pp. 565–589.
[39] Ernst Kunz and Rolf Waldi, Regular differential forms, Contemporary Mathemat-
ics, vol. 79, American Mathematical Society, Providence, RI, 1988.
[40] L. Gaunce Lewis, Jr., J. Peter May, Mark Steinberger, and James E. McClure,
Equivariant stable homotopy theory, Lecture Notes in Mathematics, vol. 1213,
Springer-Verlag, Berlin, 1986, With contributions by J. E. McClure.
[41] Joseph Lipman, Dualizing sheaves, differentials and residues on algebraic vari-
eties, Astérisque (1984), no. 117, ii+138.
[42] , Residues and traces of differential forms via Hochschild homology, Con-
temporary Mathematics, vol. 61, American Mathematical Society, Providence, RI,
1987.
[43] , Lectures on local cohomology and duality, Local cohomology and its
applications (Guanajuato, 1999), Lecture Notes in Pure and Appl. Math., vol. 226,
Dekker, New York, 2002, pp. 39–89.
[44] , Notes on derived categories and Grothendieck duality, LNM, Springer-
Verlag, to appear.
[45] Joseph Lipman, Suresh Nayak, and Pramathanath Sastry, Pseudofunctorial behav-
ior of Cousin complexes on formal schemes, Variance and duality for Cousin
complexes on formal schemes, Contemp. Math., vol. 375, Amer. Math. Soc.,
Providence, RI, 2005, pp. 3–133.
[46] Joseph Lipman and Amnon Neeman, Quasi-perfect scheme maps and bound-
edness of the twisted inverse image functor, Illinois J. Math. 51 (2007), 209–
236.
[47] Ralf Meyer and Ryszard Nest, The Baum-Connes conjecture via localisation of
categories, Topology 45 (2006), no. 2, 209–259.
[48] Daniel S. Murfet, The mock homotopy category of projectives and Grothendieck
duality, (PhD thesis, Australian National U. 2008).
[49] Suresh Nayak, Pasting pseudofunctors, Variance and duality for Cousin complexes
on formal schemes, Contemp. Math., vol. 375, Amer. Math. Soc., Providence, RI,
2005, pp. 195–271.
[50] Amnon Neeman, The connection between the K–theory localisation theorem of
Thomason, Trobaugh and Yao, and the smashing subcategories of Bousfield and
Ravenel, Ann. Sci. École Normale Supérieure 25 (1992), 547–566.
[51] , The Grothendieck duality theorem via Bousfield’s techniques and Brown
representability, Jour. Amer. Math. Soc. 9 (1996), 205–236.
[52] , Triangulated Categories, Annals of Mathematics Studies, vol. 148,
Princeton University Press, Princeton, NJ, 2001.
[53] , The homotopy category of flat modules, and Grothendieck duality, Invent.
Math. 174 (2008), 255–308.
[54] , Some adjoints in homotopy categories, Annals of Mathematics Studies,
vol. 171, Princeton University Press, Princeton, NJ, 2010, pp. 2143–2155.
[55] Dieter Puppe, On the structure of stable homotopy theory, Colloquium on algebraic
topology, Aarhus Universitet Matematisk Institut, 1962, pp. 65–71.
[56] Jean-Pierre Ramis, Gabriel Ruget, and Jean-Louis Verdier, Dualité relative en
géométrie analytique complexe, Invent. Math. 13 (1971), 261–283.
[57] Neantro Saavedra Rivano, Catégories Tannakiennes, Springer-Verlag, Berlin,
1972, Lecture Notes in Mathematics, Vol. 265.
[58] Pramathanath Sastry, Base change and Grothendieck duality for Cohen-Macaulay
maps, Compos. Math. 140 (2004), no. 3, 729–777.
[59] , Duality for Cousin complexes, Variance and duality for Cousin complexes
on formal schemes, Contemp. Math., vol. 375, Amer. Math. Soc., Providence, RI,
2005, pp. 137–192.
[60] Jean-Pierre Serre, Un théorème de dualité, Comment. Math. Helv. 29 (1955),
9–26.
[61] Nicolas Spaltenstein, Resolutions of unbounded complexes, Compositio Math. 65
(1988), no. 2, 121–154.
[62] Robert W. Thomason and Thomas F. Trobaugh, Higher algebraic K–theory of
schemes and of derived categories, The Grothendieck Festschrift (a collection of
350 Amnon Neeman
papers to honor Grothendieck’s 60’th birthday), vol. 3, Birkhäuser, 1990, pp. 247–
435.
[63] Michel van den Bergh, Existence theorems for dualizing complexes over non-
commutative graded and filtered rings, J. Algebra 195 (1997), no. 2, 662–679.
[64] Jean-Louis Verdier, Le théorème de dualité de Poincaré, C. R. Acad. Sci. Paris
256 (1963), 2084–2086.
[65] , A duality theorem in the etale cohomology of schemes, Proc. Conf. Local
Fields (Driebergen, 1966), Springer, Berlin, 1967, pp. 184–198.
[66] Jean-Louis Verdier, Base change for twisted inverse images of coherent sheaves,
vol. Collection: Algebraic Geometry, Tata Inst. Fund. Res., 1968, pp. 393–408.
[67] , Des catégories dérivées des catégories abeliennes, Asterisque, vol. 239,
Société Mathématique de France, 1996 (French).
[68] Amnon Yekutieli, Dualizing complexes over noncommutative graded algebras, J.
Algebra 153 (1992), no. 1, 41–84.
[69] , An explicit construction of the Grothendieck residue complex, Astérisque
(1992), no. 208, 127, With an appendix by Pramathanath Sastry.
[70] Amnon Yekutieli and James J. Zhang, Dualizing complexes and perverse modules
over differential algebras, Compos. Math. 141 (2005), no. 3, 620–654.
[71] , Dualizing complexes and perverse sheaves on noncommutative ringed
schemes, Selecta Math. (N.S.) 12 (2006), no. 1, 137–177.
Centre for Mathematics and its Applications, Mathematical Sciences Institute,

John Dedman Building, The Australian National University, Canberra, ACT 0200,
AUSTRALIA
E-mail address:

Derived categories and algebraic geometry
rapha ël rouquier
Contents
1. Introduction 351
2. Notations 352
3. Localisation 352
3.1. Abelian case 352
3.2. Triangulated case 354
3.2.1. Derived functors 354
3.2.2. Open subvarieties and quotients 354
3.2.3. Perfect complexes 355
3.2.4. Extensions of perfect complexes 356
3.2.5. Applications to K-theory 359
4. Reconstruction 360
4.1. Abelian case 360
4.1.1. Classification of Serre subcategories 360
4.1.2. Centres 361
4.2. Triangulated case 362
4.2.1. Classification of thick subcategories 362
4.2.2. Centres 363
4.2.3. Remarks on centres 364
4.2.4. Affine varieties 366
4.2.5. (Anti-)ample canonical bundles 366
References 369
1. Introduction
We describe some basic properties of the derived category of coherent sheaves
on a variety (bounded derived category or perfect complexes).
The first chapter considers the problem of extending vector bundles from an
open subset. Thomason and Trobaugh provided an answer to this problem by
351
352 Raphaël Rouquier
considering extensions for perfect complexes. This has applications to higher

K-theory.
In the second chapter, we explain how to characterize subcategories cor-
responding to objects supported by a given closed subvariety. This permits
a reconstruction of the variety (viewed as a ringed space) from a categori-
cal structure. In the case of derived categories, this requires also the tensor
structure.
We start with the classical case of the category of coherent sheaves (after
Gabriel). We present afterwards a similar approach in the triangulated case,
where serious difficulties arise.
Finally, we explain how to deduce that a smooth projective variety with
ample or anti-ample canonical bundle is determined by its derived category.
We haven’t included proofs of the results on general properties of abelian or
triangulated categories (cf [KaScha, Nee3] for proofs). The only difficult part
is Lemma 3.9 on compact objects.
This text is based on lectures at the conference Géométrie algébrique com-
plexe, CIRM, Luminy in December 2003. I thank Paul Balmer for useful
comments.
2. Notations
We fix a field k and we call variety a separated scheme of finite type over k.
Given a variety X, we denote by X-coh (resp. X-qcoh) the category of coherent
(resp. quasi-coherent) sheaves on X.
All functors between triangulated categories are assumed to be triangulated.
We denote by Z(C) the centre of a category C (=endomorphisms of the
identity functor).
Given a ring R, we denote by R-mod the category of finitely generated
R-modules.
3. Localisation
3.1. Abelian case
Let us recall some basic properties of categories of coherent sheaves.
Let X be a variety. Given Z closed in X, we denote by X-cohZ the full
subcategory of X-coh of coherent sheaves with support contained in Z. This is
a Serre subcategory of X-coh.
Let us recall that a full subcategory I of an abelian category A is a Serre

subcategory if it is stable under taking subobjects, quotients and extensions.
Given such a subcategory, there is a quotient abelian category A/I and a
functor A → A/I with kernel I. It is the solution of the universal problem of
taking quotients (for abelian categories). We say that there is an exact sequence
of abelian categories 0 → I → A → A/I → 0.
Let j : U = X − Z → X be the open embedding.
Proposition 3.1. The functor j ∗ : X-coh → U -coh induces an equivalence

∼
X-coh /X-cohZ → U -coh, i.e., there is an exact sequence of abelian categories
0 → X-cohZ → X-coh → U -coh → 0.
Proof. In the case of quasi-coherent sheaves, we have a functor j∗ right adjoint

∼
to j ∗ . The canonical map j ∗ j∗ → 1U -qcoh is an isomorphism and the kernel of
j ∗ is X-qcohZ . It follows from Lemma 3.2 below that there is an exact sequence
0 → X-qcohZ → X-qcoh → U -qcoh → 0.
Let us now deduce the proposition from the characterisation of coherent

sheaves as the finitely presented objects in the category of quasi-coherent
sheaves. Lemma 3.3 below shows that the canonical functor X-coh /X-cohZ →
U -coh is fully faithful. We are left with proving that the functor j ∗ : X-coh →
U -coh is essentially surjective. Let G be a coherent sheaf on U and let F =
j∗ G. We have j ∗ F G. The quasi-coherent sheaf F is an increasing union

(filtered colimit) of its coherent subsheaves, F = E coherent ⊂F E. It follows

that j ∗ F = E j ∗ E. Since j ∗ F is coherent and it is an increasing union of a
family of subsheaves, one of the members of the family j ∗ E is equal to j ∗ F
(the filtered colimit stabilizes after finitely many terms). So, j ∗ F = j ∗ E G
for a coherent subsheaf E of F .
Lemma 3.2. Let F : A → B be an exact functor between abelian categories.

can
Assume F has a right adjoint G and G is fully faithful (i.e., F G −→ 1B is an
isomorphism).
Then ker F is a Serre subcategory of A and there is an exact sequence
0 → ker F → A → B → 0.
Lemma 3.3. Let A be an abelian category, A a full abelian subcategory of

A and I a Serre subcategory of A. Assume that for any M ∈ A and for any
N ∈ I a subobject or a quotient of M, then N ∈ A .
Then the canonical functor A /(I ∩ A ) → A/I is fully faithful.
3.2. Triangulated case

3.2.1. Derived functors
We start by recalling some properties of derived categories of sheaves and we
explain how to construct right derived functors.
Recall that the canonical functor D(X-coh) → D(X-qcoh) is fully faithful,
i.e., D(X-coh) is equivalent to the full subcategory of D(X-qcoh) of complexes
whose cohomology sheaves are coherent. We will identify those two categories.
Let X-inj be the category of quasi-coherent injective sheaves and Ho(X-inj)
the homotopy category of complexes of objects of X-inj. Consider the canonical
functor Ho(X-qcoh) → D(X-qcoh). It has a right adjoint ρ (“homotopically
injective resolution”). Let Ho(X-qcoh)hi be its essential image (homotopically
injective complexes). The functor ρ is fully faithful, the canonical functor
∼
Ho(X-qcoh)hi → D(X-qcoh) is an equivalence with inverse ρ. The intersec-
tion of Ho(X-qcoh)hi with D + (X-qcoh) is Ho+ (X-inj), i.e., ρ restricts to an
∼
equivalence D + (X-qcoh) → Ho+ (X-inj) : we recover the classical injective
resolutions.
Let us now discuss the derivation of a left exact functor F : X-qcoh →
A, where A is an abelian category. We extend F to a functor Ho(F ) :
Ho(X-qcoh) → Ho(A). We restrict this functor to Ho(X-qcoh)hi . We obtain
the right derived functor
ρ Ho(F ) can
RF : D(X-qcoh) −
→ Ho(X-qcoh)hi −−−→ Ho(A) −→ D(A).
∼
The functor RF is triangulated. In particular, the image of a distinguished

triangle is a distinguished triangle, while F needs not send an exact sequence
∼
to an exact sequence. The left exactness of F shows that H 0 (RF (M)) → F (M)
for M ∈ X-qcoh.
3.2.2. Open subvarieties and quotients

The functor j∗ derives into a functor Rj∗ : D(U -qcoh) → D(X-qcoh). This is
right adjoint to the functor j ∗ : D(X-qcoh) → D(U -qcoh). The kernel of the
functor j ∗ is DZ (X-qcoh), the full subcategory of D(X-qcoh) of complexes
whose cohomology sheaves have their support contained in Z. This is a thick
subcategory.
Let us recall that a non-empty full subcategory I of a triangulated category
T is thick if the following two conditions hold
r Given F → G → H a distinguished triangle in T , if two objects amongst
F , G and H are in I, then the third one is in I as well.
r Given F, G ∈ T , if F ⊕ G ∈ I, then F, G ∈ I.
There is a quotient triangulated category T /I and a functor T → T /I with

kernel I, solution of the universal quotient problem (amongst triangulated
categories). We say that there is an exact sequence of triangulated categories
0 → I → T → T /I → 0.
Lemma 3.4 gives an exact sequence of triangulated categories
0 → DZ (X-qcoh) → D(X-qcoh) → D(U -qcoh) → 0
Lemma 3.4. Let F : T → T be a functor between triangulated categories.

Assume F has a right adjoint G and G is fully faithful.
Then ker F is a thick subcategory of T and there is an exact sequence
0 → ker F → T → T → 0.
3.2.3. Perfect complexes

We come now to the core of our study. We recall the notion of perfect objects
and their basic properties.
An object of D(X-qcoh) is perfect if it is locally (quasi)-isomorphic to a
bounded complex of free sheaves of finite rank. We denote by X-perf the full
subcagegory of D(X-qcoh) of perfect complexes. This is a thick subcategory
of D b (X-coh). If X is quasi-projective, then a complex is perfect if and only if
it is quasi-isomorphic to a bounded complex of vector bundles. The variety X
is regular if and only if D b (X-coh) = X-perf.
Let T be a triangulated category with infinite direct sums. An object
C ∈ T is compact if given any family E of objects of T , the canonical map

E∈E Hom(C, E) → Hom(C, E∈E E) is an isomorphism. We denote by T
c
the full subcategory of T of compact objects. This is a thick subcategory.

A key idea in Thomason’s approach is the characterisation of perfect com-
plexes as the compact objects of D(X-qcoh) (cf [Rou2] for a study of compact-
ness as a local property for general triangulated categories).
Lemma 3.5. Let C ∈ D(X-qcoh). Then C is perfect if and only if it is compact.
Proof. Let us assume first X is affine, X = Spec R. Since R is compact, we

deduce that every perfect object is compact. Let C be a complex of R-modules
and let i ∈ Z such that H i C = 0. Then Hom(R, C[i]) = 0. It follows from
Lemma 3.9 that the perfect complexes are the same as the compact objects.
We consider now an arbitrary variety X. Let X = U1 ∪ U2 with U1 an
affine open subvariety and U2 open. We assume that the minimal number of
open affine subvarieties in a covering of U2 is strictly less that the number
for X. By induction, we can assume the lemma holds for U2 and for U12 =
U1 ∩ U2 . Let jr : Ur → X and j12 : U12 → X be the open immersions. Given
D ∈ D(X-qcoh), there is a Mayer-Vietoris distinguished triangle:

D → Rj1∗ j1∗ D ⊕ Rj2∗ j2∗ D → Rj12∗ j12
∗
D
Let C ∈ D(X-qcoh). The triangle above shows that C is compact if j1∗ C, j2∗ C
∗
and j12 C are compact. The converse is clear. On the other hand, C is perfect if
and only if j1∗ C, j2∗ C and j12
∗
C are perfect. The lemma follows by induction.
An important aspect of the category X-perf is that it provides the “right”
K-theory groups, for a variety without enough ample vector bundles. We put
K0 (X) = K0 (X-perf).
Recall that given a triangulated category T , we define K0 (T ) as the quotient
of the free abelian group with basis the isomorphism classes of objects of T
by the relation [M] = [L] + [N ] whenever there is a distinguished triangle
L → M → N .
This definition of K0 (X) coincides with the classical one (Grothendieck
group of the exact category of vector bundles) when X has an ample family of
line bundles (this is the case for a quasi-projective variety).
3.2.4. Extensions of perfect complexes

We put X-perf Z = X-perf ∩DZ (X-qcoh).
Theorem 3.6 (Thomason-Trobaugh). The functor j ∗ induces a fully faithful
functor
X-perf /X-perf Z → U -perf .
An object of U -perf is the restriction of an object of X-perf if and only if its
class in K0 (U ) is the restriction of an element of K0 (X).
The occurrence of K0 in Theorem 3.6 comes from the following lemma of
Thomason [Th2, Theorem 2.1] (the proof is tricky).
Lemma 3.7. Let T be a triangulated category. There is a bijection from the set
of full triangulated subcategories I of T that generate T as a thick subcategory
to the set of subgroups of K0 (T ) given by sending I to the image of K0 (I) in
K0 (T ).
The calculus of fractions gives a simple criterion for fully faithfulness:
Lemma 3.8. Let T be a triangulated category, I a thick subcategory of T and
T a full triangulated subcategory of T .
Assume every morphism C → D with C ∈ T and D ∈ I factors through
an object of I ∩ T . Then the canonical functor T /(I ∩ T ) → T /I is fully
faithful.
Given I a full subcategory of a triangulated category T , we denote by Ī the

smallest thick subcategory of T closed under taking infinite direct sums and
containing I.
Let I be a thick subcategory of a triangulated category T . The right orthogo-
nal I ⊥ to I in T is the full subcategory of T of objects D with Hom(C, D) = 0
for all C ∈ I. This is a thick subcategory.
The following lemma is related to Brown-Neeman’s representability Theo-
rem and its proof requires some work [Nee2].
Lemma 3.9. Let T be a triangulated category with arbitrary direct sums.

Let I be a thick subcategory of T c . Then every map from an object of T c
to an object of Ī factors through an object of I. In particular, we have
T c ∩ Ī = I.
We have Ī = T if and only if the right orthogonal I ⊥ of I in T vanishes.
Lemma 3.10. Let Y be a closed subvariety of X. We have DY (X-qcoh) =

X-perf Y .
Given Z closed in X, we put K0 (X on Z ) = K0 (X-perf Z ). We will show

a version “with supports” of Theorem 3.6:
Theorem 3.11. Let Z be a closed subvariety of X. The functor j ∗ induces

a fully faithful functor X-perf Z /X-perf Z∩Z → U -perf U ∩Z . An object of
U -perf U ∩Z is the image of an object of X-perf Z if and only if its class in
K0 (U on U ∩ Z ) is the image of an element of K0 (X on Z ).
Proof of Theorem 3.11 and Lemma 3.10. Let us show first that the lemma for X
implies the theorem for X. The combination of Lemmas 3.8, 3.9 and 3.10 shows
that j ∗ induces a fully faithful functor X-perf Z /X-perf Z∩Z → U -perf U ∩Z .
Let I be the image of that functor: this is a full triangulated subcategory.
Since X-perf Z = DZ (X-qcoh) (lemma 3.10), we have Ī = DU ∩Z (U -qcoh).
It follows from Lemma 3.9 that U -perf Z ∩U is the thick subcategory generated
by I. The theorem follows now from Lemma 3.7.
Lemma 3.9 shows that Lemma 3.10 will follow from the fact that
(X-perf Y )⊥ = 0.
Let us assume first that X is affine. Let {y1 , . . . , yr } be a family of generators
of the defining ideal of Y and let
6
r
yi
Gr = (0 → OX −
→ OX → 0)
i=1
be the associated Koszul complex (the non-zero terms are in degrees −r, . . . , 0).
We will show by induction on r that an object C ∈ DY (X-qcoh) vanishes if
Gr ⊗ C = 0. The lemma will follow, since Hom(G∨r , C[i]) H i (Gr ⊗ C),
where G∨r = R Hom(Gr , OX ) is the dual of Gr . The case r = 0 is clear.
Consider r > 0 and C ∈ DY (X-qcoh) a non-zero object. By induction, there
exists i such that H i (Gr−1 ⊗ C) = 0. The distinguished triangle
yr
Gr−1 ⊗ C −
→ Gr−1 ⊗ C → Gr ⊗ C
yr
gives an exact sequence H i−1 (Gr ⊗ C) → H i (Gr−1 ⊗ C) − → H i (Gr−1 ⊗ C).
Since H (Gr−1 ⊗ C) is supported by the closed subvariety (yr = 0), we deduce
i
that multiplication by yr has a non-zero kernel, hence H i−1 (Gr ⊗ C) = 0. This

completes the proof of Lemma 3.10 in the affine case.
We will now prove the lemma by induction on the minimal number of open
affine subsets in a covering of X.
Let X = U1 ∪ U2 with U1 open affine and U2 open for which the lemma
holds. We put Zi = X − Ui . Let C ∈ DY (X-qcoh) with Hom(D, C) = 0 for
all D ∈ X-perf Y .
Let D ∈ U1 -perf Y ∩Z2 . The functor Rj1∗ : D(U1 -qcoh) → D(X-qcoh)
∼
restricts to equivalences DY ∩Z2 (U1 -qcoh) → DY ∩Z2 (X-qcoh) and
∼
U1 -perf Y ∩Z2 → X-perf Y ∩Z2 . It follows that Hom(Rj1∗ D, C) = 0. Let C
can
be the cocone of the adjunction morphism C −→ Rj2∗ j2∗ C. We have
Hom(Rj1∗ D, Rj2∗ j2∗ C[n]) Hom(j2∗ Rj1∗ D, j2∗ C[n]) = 0
for all n, hence Hom(Rj1∗ D, C ) Hom(Rj1∗ D, C) = 0. Since C is supported

by Y ∩ Z2 , there exists C ∈ DY ∩Z2 (U1 -qcoh) such that C = Rj1∗ C . We have
Hom(D, C ) Hom(Rj1∗ D, C ) = 0. The affine case of the lemma shows that
C = 0, hence C Rj2∗ j2∗ C.
Let E ∈ U2 -perf Y ∩U2 , E = E ⊕ E [1] and G = E|U1 ∩U2 . The affine case
of the theorem shows that there exists F ∈ U1 -perf Y ∩U1 and an isomorphism
∼
F|U1 ∩U2 → G. Let now D be the cocone of the morphism sum of the adjunc-
tion morphisms Rj2∗ E ⊕ Rj1∗ F → Rj12∗ G. Then j1∗ D F and j2∗ D E,
hence D ∈ X-perf Y . We have Hom(E, j2∗ C) Hom(D, Rj2∗ j2∗ C) = 0. By
induction, we deduce that j2∗ C = 0, hence C = 0. This proves Lemma 3.10
for X.
Exercice 3.1. Let C ∈ X-qcoh such that for any set I of objects of X-qcoh, the

canonical map D∈I Hom(C, D) → Hom(C, D∈I D) is an isomorphism.
Show that C is coherent.
A striking special case of Theorem 3.6 is given by the following corollary.
Corollary 3.12. Let L be a vector bundle on U . Then there exists a perfect

complex on X whose restriction to U is quasi-isomorphic to L ⊕ L[1].
Remark 3.13. Let us show, following Serre [Se, 5.a, p.371], that Theorem 3.6
doesn’t hold for vector bundles.
Let X = A3 and U = X − {0}. Let F be the vector bundle on U which is
the pullback of the tangent bundle on P2 . The restriction map K0 (X) → K0 (U )
is an isomorphism. Since F is not the direct sum of two line bundles, it is not
the restriction of a vector bundle on X.
Let G be a coherent sheaf on X extending F. The second syzygy #2 G of G
is locally free (hence free) and this provides a perfect complex extending F :
there is a complex of free sheaves
0 → #2 G → P −1 → P 0 → 0
with homology concentrated in degree 0 and isomorphic to G.
Remark 3.14. A proof similar to that of Proposition 3.1 shows that there is
an exact sequence 0 → DZb (X-coh) → D b (X-coh) → D b (U -coh) → 0. When
X is smooth, then X-perf = D b (X-coh), hence Theorem 3.6 is a consequence
of that exact sequence. In this case, the canonical map K0 (X) → K0 (U ) is
surjective.
3.2.5. Applications to K-theory

From Theorem 3.6, Thomason deduces a long exact sequence for higher K-
theory, via Waldhausen’s theory [Th1]:
Theorem 3.15. There is a long exact sequence
· · · → Ki (X on Z) → Ki (X) → Ki (U ) → Ki−1 (X on Z) → · · ·
Thomason deduces also an excision result.
Theorem 3.16. If X = U ∪ V with V open and Z ⊂ V , then there are iso-

∼
morphisms Ki (X on Z) → Ki (V on Z).
Finally, he obtains a Mayer-Vietoris Theorem.
Theorem 3.17. Let U and V be open subsets of X. Then there is a long exact
sequence
· · · → Ki (U ∪ V ) → Ki (U ) ⊕ Ki (V ) → Ki (U ∩ V ) → Ki−1 (U ∪ V ) → · · ·
These sequences can be extended to negative i, via a version of Bass’

fundamental Theorem:
Theorem 3.18. There is an exact sequence
0 → Ki (X) → Ki (X[T ]) ⊕ Ki (X[T −1 ]) → Ki (X[T , T −1 ]) → Ki−1 (X) → 0
Classical methods based on exact categories had led to similar results under
restrictive assumptions. Thomason obtains also a local-global principle for
Ki ’s.
4. Reconstruction
4.1. Abelian case
We will start with the classical case of coherent sheaves, following Gabriel
[Ga].
4.1.1. Classification of Serre subcategories

We say that a Serre subcategory I of an abelian category A is of finite type if it
is generated by an object (i.e., the smallest Serre subcategory of A containing
the object is I). We say that a Serre subcategory I is irreducible if it is not
equal to 0 and if it is not generated by two proper Serre subcategories of I.
Theorem 4.1 (Gabriel). The map Z
→ X-cohZ from the set of closed subsets
of X to the set of Serre subcategories of finite type of X-coh is a bijection.
The closed irreducible subsets correspond to the irreducible Serre subcate-
gories.
This follows immediately from the next lemma:
Lemma 4.2. A coherent sheaf with support Z generates X-cohZ as a Serre
subcategory.
Proof. Let F be a coherent sheaf with support Z and let I be the Serre subcat-
egory of X-coh generated by F . Let i : Y → X be the closed embedding of a
subvariety. Every coherent sheaf on X supported by Y is an extension of sheaves
of the form i∗ G. Let J be the Serre subcategory of Y -coh generated by i ∗ F .
Since i∗ i ∗ F ∈ I, we have i∗ (J ) ⊂ I. If J = Y -cohY ∩Z , then X-cohY ∩Z ⊂ I.
If in addition Z ⊂ Y , then I = X-cohZ .
It follows that it is enough to prove the lemma for X reduced and Z = X.
We proceed by induction on the dimension n of X, then on the number of
irreducible components of dimension n of X, then on the number of irreducible
components of dimension n − 1 of X, etc.
Let Y be a proper closed subset of X. The discussion above shows that by

induction we can assume X-cohY ⊂ I.
Let M be a coherent sheaf on X. Let U be an irreducible open affine subset of
X and j : U → X the open immersion. Shrinking U if necessary, the sheaves
∼
j ∗ M and j ∗ F are free, of respective ranks r and s > 0. Let f : j ∗ F r → j ∗ M s

be an isomorphism. By Proposition 3.1, there is a coherent sheaf M on X, there
are ψ : F r → M and φ : M s → M such that j ∗ (ψ) = j ∗ (φ)f and j ∗ (φ) is
an isomorphism. The kernels and cokernels of φ and ψ have their supports
contained in X − U , hence they are in I. It follows that M ∈ I.
4.1.2. Centres
Lemma 4.3. Let R be a ring. The canonical map Z(R) → Z(R-mod) is an
isomorphism.
Proof. Evaluation at R gives a left inverse. Let α ∈ Z(R-mod) with α(R) = 0.

Since any R-module is a quotient of a free R-module, it follows that α = 0.
Corollary 4.4 (Gabriel). The abelian category X-coh determines the variety
X.
Proof. We define a ringed space E. Its points are the irreducible Serre sub-
categories of finite type of X-coh. The open subsets are the D(I), defined
as the set of those subcategories J that are not contained in a given Serre
subcategory I.
Theorem 4.1 shows that the map sending a point x ∈ X to X-coh{x} defines
a homeomorphism X → E.
Consider the presheaf of rings on E given by OE (D(I)) = Z(X-coh /I).
If D(I ) ⊂ D(I), then the quotient functor X-coh /I → X-coh /I induces a
map Z(X-coh /I) → Z(X-coh /I ). We denote by OE the associated sheaf.
The canonical map (U ) → Z(U -coh) induces a morphism of ringed spaces
X → E. To check that this is an isomorphism, it is enough to consider its
restriction to an open affine subset: Lemma 4.3 provides the conclusion in the
affine case.
Remark 4.5. Actually, Lemma 4.3 holds for non-affine varieties: the canonical
map (OX ) → Z(X-coh) is an isomorphism of rings. That follows from the
construction of the category X-coh by gluing the abelian categories Ui -coh
along the quotient categories (Ui ∩ Uj )-coh, given a finite open covering of X
by open affine subsets Ui .
As a consequence, the presheaf in the proof of Corollary 4.4 is a sheaf.
4.2. Triangulated case

4.2.1. Classification of thick subcategories
Inspired by work on the stable homotopy category (description of the chromatic
tower), Hopkins and Neeman [Ho, Nee1] have given a classification of thick
subcategories of the category of perfect complexes over an affine variety. This
result was generalised later by Thomason [Th2].
Let I be a thick subcategory of X-perf. We say that I is of finite type if it
is generated by an object (i.e., if I is the smallest thick subcategory of X-perf
containing the object). We say that I is an ideal if it is thick and if given any
C ∈ I and D ∈ X-perf then C ⊗L D ∈ I. We say that an ideal I is irreducible
if it is non-zero and it is not generated by two proper ideals.
Theorem 4.6 (Hopkins, Neeman, Thomason). The map Z

→ X-perf Z from
the set of closed subsets of X to the set of ideals of finite type of X-perf is a
bijection.
Irreducible closed subsets correspond to irreducible subcategories.
The starting point is the following Lemma.
Lemma 4.7. Let Z be a closed subset of X. Then there exists a perfect complex
on X with support Z.
Proof. Assume first Z is irreducible and X = Spec R is affine. Consider equa-

7 fi
tions f1 = 0, . . . , fn = 0 defining Z. The support of i (0 → R − → R → 0)
is Z: this solves the lemma in that case.
We assume now that Z is irreducible but X is not necessarily affine. Let U
be an open affine subset of X containing the generic point of Z. Then there
exists C ∈ U -perf with support U ∩ Z. The version “with supports” of the
localisation theorem (Theorem 3.11) shows that there exists D ∈ X-perf Z such
that D|U C ⊕ C[1]. We have Supp(D) ∩ U = Z ∩ U , hence Supp(D) = Z.
We consider finally the general case. Let Z = Z1 ∪ · · · ∪ Zr be the decom-
position in irreducible components and consider Ci ∈ X-perf with support Zi .

Then the support of Ci is Z.
Theorem 4.6 follows now from the next lemma.
Lemma 4.8. A perfect complex on X with support Z generates X-perf Z as an

ideal.
Proof. Let C ∈ X-perf with support Z and let I be the ideal of X-perf generated
by C. Lemmas 3.9 and 3.10 show that I = X-perf Z if and only if DZb (X-coh) ⊂
Ī.
Given a closed immersion i : Y → X, we have i∗ Li ∗ C C ⊗L OY ∈

Ī since OY ∈ D(X-qcoh) = X-perf (lemma 3.9). In addition, Li ∗ C ∈
Y -perf Y ∩Z . Every object of DYb ∩Z (X-coh) is a finite extension (=iterated cone)
of objects i∗ G with G ∈ DYb ∩Z (Y -coh). Let J be the ideal of Y -perf Y ∩Z gen-
erated by Li ∗ C. Since i∗ (Li ∗ C ⊗L M) C ⊗L i∗ M for all M ∈ D(Y -qcoh),
we have i∗ (J¯ ) ⊂ Ī. If J¯ = DY ∩Z (Y -qcoh), then DYb ∩Z (X-coh) ⊂ i∗ (J¯ ). If in
addition Z ⊂ Y , then I = X-perf Z .
It follows that it is enough to prove the lemma for X reduced and Z = X.
We proceed by induction on the dimension n of X, then on the number of
irreducible components of dimension n of X, then on the number of irreducible
components of dimension n − 1 of X, etc.
Let Y be a proper closed subset of X. The discussion above shows that
X-perf Y ⊂ I.
Let M ∈ X-perf. There is a non-empty open affine subset j : U → X
such that j ∗ M and j ∗ C are finite sums of complexes OU [r]. So, there are
finite-dimensional graded vector spaces (viewed as complexes with vanishing
∼
differential) V = 0 and W and there is an isomorphism f : j ∗ (C ⊗k W ) →
j ∗ (M ⊗k V ). Theorem 3.6 shows that there is an object M ∈ X-perf, maps
ψ : C ⊗k W → M and φ : M ⊗k V → M such that j ∗ (ψ) = j ∗ (φ)f and
j ∗ (φ) is an isomorphism. The cones of φ and ψ have a support contained in
the closed subset X − U of X, so they are in I by induction. It follows that
M ∈ I.
Remark 4.9. The classical proof of Theorem 4.6 uses the following result
(“tensor nilpotence Theorem”).
Let C ∈ X-perf, let D ∈ D(X-qcoh) and let f : C → D. We assume that
for every point x of X, we have f ⊗ k(x) = 0 in D(k(x)-Mod).
Then there is an integer n such that ⊗n f : ⊗n C → ⊗n D vanishes in
D(X-qcoh).
4.2.2. Centres
We proceed as in §4.1.2 to obtain a reconstruction Theorem [Ba1, Rou1].
Let X be a variety. We define Z(X-perf)lnil as the subring of Z(X-perf)
given by elements α such that α(C) is nilpotent for all C ∈ X-perf. We put
Z(X-perf)lred = Z(X-perf)/Z(X-perf)lnil .
Lemma 4.10. The canonical morphism (OX ) → Z(X-perf) induces an iso-

∼
morphism (OX )red → Z(X-perf)lred .
Proof. Evaluation at OX gives a left inverse to the canonical map (OX ) →

Z(X-perf).
Assume first X = Spec R is affine. Let α ∈ Z(R-perf) such that α(R) = 0.

A perfect complex is quasi-isomorphic to a bounded complex C of finitely
generated projective R-modules. Let n = max{i|C i = 0} − min{i|C i = 0}. By
induction on n, one sees that α(C)n+1 = 0 and the lemma follows in the affine
case.
Consider now an arbitrary variety X and α ∈ Z(X-perf) such that α(OX ) =
0. Let U be an open affine subset of X and let αU ∈ Z(U -perf) be the element
induced by α. We have αU (OU ) = 0, hence for any C ∈ U -perf, the endomor-
phism αU (C) is nilpotent. Let X = U1 ∪ · · · ∪ Ur be a covering by open affine
subsets and let V = U2 ∪ · · · ∪ Ur . We prove the lemma by induction on r. Let
C ∈ X-perf and n > 0 such that αU1 (C|U1 )n = 0. Then α(C)n factors through
an object C ∈ X-perf Z , where Z = X − U1 and α(C)(d+1)n factors through
α(C )dn for all d ≥ 0. Since Z ⊂ V , the restriction functor X-perf Z → V -perf Z

is fully faithful. By induction, αV (C|V ) is nilpotent, hence α(C ) is nilpotent
and α(C) is nilpotent as well.
Theorem 4.11 (Balmer, R.). If X is a reduced variety, then the category X-perf,
viewed as a tensor triangulated category, determines X.
It would be more satisfactory not to use the tensor structure. Unfortunately,
there are too many thick subcategories in general. Balmer has developed a
general approach to the study of the geometry of tensor triangulated categories
[Ba2, BaFl].
Example 4.12. Let X = P1 . Then the thick subcategory of X-perf generated
by O is equivalent to D b (k-mod). It is not of the form X-perf Z .
4.2.3. Remarks on centres

Let us discuss in more details centres of categories of perfect complexes. Note
that the difficulties are due to the weakness of the axioms of triangulated cat-
egories and can be solved using dg-categories [To]. The next two Propositions
show that the centre of the category of perfect complexes has no non-zero
nilpotent elements in some cases.
Proposition 4.13. Let R be a ring without zero divisors. Then the canonical
map Z(R) → Z(R-perf) is an isomorphism.
Proof. Evaluation at R defines a morphism α : Z(R-perf) → Z(R). The canon-
ical map Z(R) → Z(R-perf) is a right inverse. We will show that α is injective.
Let z ∈ Z(R-perf) with z(R) = 0. Let C be a non-zero bounded complex of
finitely generated projective R-modules. Consider r minimal such that C r = 0
and s maximal such that C s = 0. We will show by induction on s − r that
z(C) = 0. This is clear when s = r. Assume s > r.
Since C r is a finitely generated projective module, there is a finitely gen-

erated projective module P such that C r ⊕ P is a free module of finite rank.
id
Let D = C ⊕ (0 → P −
→ P → 0), where the non-zero terms of the complex
id
0→P − → P → 0 are in degrees r and r + 1. Then D r is free of finite rank
and C is homotopy equivalent to D. So, it is enough to prove that z(C) = 0
when C r is free.
We proceed now by induction on the rank of C r . Let C r = L1 ⊕ · · · ⊕ Ln
be a decomposition into free modules of rank 1. Consider an integer i with
1 ≤ i ≤ n. Let D be the subcomplex of C given by D l = C l pour l = r and
D r = L1 ⊕ · · · ⊕ Li−1 ⊕ Li+1 ⊕ · · · ⊕ Ln . By induction, we have z(D) = 0. It
can z(C)
follows that the composition f : D −→ C −−→ C is homotopic to 0. Consider
{hl : D l → C l−1 }l with f = dC h + hdD . The morphism h extends uniquely
into a graded endomorphism k of degree −1 of C that vanishes on Li . Let
can
ψ = z(C) − (dC k + kdC ), a map homotopic to z(C). The composition D −→
ψ
C− → C vanishes, hence ψ l = 0 for l = r and D r ⊂ ker ψ r . Note that im ψ r ⊂
ker dCr . If ψ r = 0 then ker dCr = 0. Since C r /D r Li is free of rank 1 and
since C r is free, if ψ r is non-zero, then its restriction to Li is injective (recall
that R has no zero-divisor) and ker ψ r = D r .
ψ can
The composition C − → C −→ C r [−r] is homotopic to 0 since z(C r [−r]) =
0. So, there is g : C r+1 → C r such that ψ r = gdCr , hence ker dCr ⊂ ker ψ r .
Assume z(C) = 0. Then ker dCr ⊂ L1 ⊕ · · · ⊕ Li−1 ⊕ Li+1 ⊕ · · · ⊕ Ln . This
holds for all i, hence ker dCr = 0. It follows that ψ r = 0, hence ψ = 0 and
finally z(C) = 0.
Proposition 4.14. Let X be an irreducible reduced quasi-projective variety.

Then, the canonical morphism (X, OX ) → Z(X-perf) is an isomorphism.
Proof. Every object of X-perf is isomorphic to a bounded complex C whose

terms are direct sums of line bundles. The proof is then the same as that of
Proposition 4.13 with free modules of rank 1 replaced by line bundles.
Remark 4.15. Let k be a field and let C be a k-linear category with finite
dimensional Hom-spaces. Let P be an indecomposable object of C and let
φ ∈ Z(End(P )) with the following properties:
r φ2 = 0
r given x ∈ End(P ) such that φx = 0 or xφ = 0, then x is invertible.
Given Q an indecomposable object of C not isomorphic to P , then
r Hom(P , Q)φ Hom(Q, P ) = 0
r for f ∈ Hom(P , Q) non zero, then, φ ∈ Hom(Q, P )f and for g ∈
Hom(Q, P ) non zero, then, φ ∈ g Hom(P , Q).
φ
Let C = (0 → P − → P → 0), a complex with non-zero terms in degrees 0
and 1. Define an endomorphism ζ of C as φ in degree 1 and 0 elsewhere.
Then, one shows there is a unique element of the centre of the homotopy
category T of complexes (all, bounded, bounded above or bounded below, . . . )
of objects of C with the following properties:
r it is 0 on indecomposable objects of T which are not isomorphic to C[i] for
some i.
r it is φ[i] on C[i].
This applies to C the category of finitely generated projective A-modules
when A = k[x]/(x 2 ) (or A = Z/4Z by slightly modifying the setting above):
the centre of A-perf is larger that A.
Remark 4.16. Proposition 4.14 does not extend to Hochschild cohomology
[Ca]. Let X be an elliptic curve. Then, H H 2 (X) = Ext2X×X (OX , OX ) =
0. On the other hand, X-coh is a hereditary category. In particular,
Hom(Id, Id[2]) = 0, where Id is the identity functor of D b (X-coh).
4.2.4. Affine varieties

Let L be an ample line bundle on X. Then X-perf is generated by the powers
L⊗−i for i > 0, as a thick subcategory (cf Lemma 3.9). It follows that a thick
subcategory I is an ideal if for any C ∈ I, we have C ⊗ L−1 ∈ I.
We deduce that if X is affine, then every thick subcategory of X-perf is an
ideal. We obtain a corollary to Theorem 4.11 (actually, Lemma 4.10 gives a
direct proof in that case).
Corollary 4.17. If X is affine and reduced, then the triangulated category
X-perf determines X.
4.2.5. (Anti-)ample canonical bundles

To study more interesting situations, let us introduce Serre functors, following
Bondal and Kapranov.
Let C be a k-linear category. A Serre functor for C is an equivalence of
∼
categories S : C → C together with the data, for every X, Y ∈ C, of bifunctorial
isomorphisms
∼
Hom(X, Y )∗ → Hom(Y, S(X)).
A Serre functor is unique up to unique isomorphism, when it exists.
Lemma 4.18. Let X be a smooth projective variety of pure dimension n. Then
S = ωX [n] ⊗ − is a Serre functor for D b (X-coh).
Proof. We can assume that X is irreducible. Given C ∈ D b (X-coh), consider

the Hom-pairing Hom(O, C) × Hom(C, ωX [n]) → H n (X, ωX ) k. When the
homology sheaves of C are concentrated in one degree, Serre’s duality Theorem
shows that this pairing is perfect. Since the thick subcategory of C’s for which
the pairing is perfect is thick, we deduce that the pairing is perfect for all C.
∼
Via the canonical isomorphisms Hom(C, D) → Hom(O, RHom(C, D))
∼
and Hom(D, C ⊗ ωX [n]) → Hom(RHom(C, D), ωX [n]), we obtain a perfect
pairing
Hom(C, D) × Hom(D, C ⊗ ωX [n]) → k.
Theorem 4.19 (Bondal-Orlov). Let X be a smooth projective variety such that

−1
ωX or ωX is ample. Then the triangulated category D b (X-coh) determines X.
If Y is a smooth projective variety, then an equivalence of triangulated
∼ ∼
categories D b (X-coh) → D b (Y -coh) gives rise to an isomorphism X → Y .
Proof. The crucial point is the fact that the Serre functor is intrinsic to the
category D b (X-coh). Since the thick subcategories invariant by the Serre func-
tor and its inverse are ideals (cf §4.2.4 and Lemma 4.18), we recover X from
D b (X-coh).
∼
Consider now F : D b (X-coh) → D b (Y -coh). Note that F commutes with
the Serre functors : F SX SY F .
Let Z be a closed subset of Y . Since F −1 (DZb (Y -coh)) is a thick subcategory
of D b (X-coh) stable under SXi for all i, it is of the form D b (Z) (X-coh) and this
provides an injection from closed subsets of Y to closed subsets of X.
Assume Z is irreducible. Let V be an open affine subset of Y . Then
∼
F −1 induces an equivalence D b (V -coh) → D b ((X − (Y − V ))-coh) that
∼
restricts to an equivalence DZb (V -coh) → D b (Z) ((X − (Y − V ))-coh). Since
DZb (V -coh) is an irreducible thick subcategory (§4.2.4 and Theorem 4.6), we
deduce that D b (Z) ((X − (Y − V ))-coh) is an irreducible thick subcategory
of D b ((X − (Y − V ))-coh), hence (Z) ∩ (X − (Y − V )) is irreducible.
If Y = V1 ∪ · · · ∪ Vr is a covering by open affine subsets, then the subsets
X − (Y − Vi ) give an open affine covering of X, and it follows that (Z) is
irreducible.
We define an injection φ : Y → X between points by φ(y) = (y). If y is
b
a closed point, then by Lemma 4.21 below the thick subcategory D{y} (Y -coh)
b
of D (Y -coh) is minimal as a non-zero thick subcategory. It follows that
F −1 (D{y}
b
(Y -coh)) = Dφ(y) (X-coh) is a minimal non-zero thick subcategory
of D (X-coh) that is stable under SXi for all i. We deduce that φ(y) is a closed
b
point.
Let x be a closed point of X that is not in the image of φ. Then

Hom(O{x} , C[i]) = 0 for all i ∈ Z and all C ∈ D{φ(y)}b
(X-coh). It follows that
Hom(F (O{x} ), O{y} [i]) = 0 for all closed points y of Y and all i ∈ Z. We deduce
from Lemma 4.20 below that F (O{x} ) = 0, a contradiction. So, φ is bijective.
Let Z be a closed subset of Y . A closed point y of Y is in Z if and only if
b
D{y} (Y -coh) ⊂ DZb (Y -coh). We have x ∈ (Z) if and only if D{x} b
(X-coh) ⊂
D (Z) (X-coh), hence φ(Z) = (Z) is closed in X. This shows that ψ = φ −1 :
b
X → Y is continuous.
Let U = Y − Z. We have a sequence of canonical isomorphisms (cf
Lemma 4.10)
∼ ∼ ∼
(U ) → Z(D b (U -coh))lred −−→ Z(D b (φ(U )-coh))lred → (φ(U )).
F −1
This extends ψ : X → Y into an isomorphism of ringed spaces.
Lemma 4.20. Let X be a variety and let C ∈ D b (X-coh) such that

Hom(C, O{x} [i]) = 0 for all closed points x in X and for all i ∈ Z. Then
C = 0.
Proof. Consider i maximal such that Hi (C) = 0 and let x be a closed
point. The canonical map Hom(Hi (C), O{x} ) → Hom(C, O{x} [−i]) is injec-
tive. Given x in the support of Hi (C), we have Hom(Hi (C), O{x} ) = 0, hence
Hom(C, O{x} [−i]) = 0.
Lemma 4.21. Let X be an algebraic variety and let x be a closed point of X.

Then X-perf {x} is a minimal non-zero thick subcategory of X-perf.
Proof. Let U be an open affine subset of X containing x. The restriction
functor X-perf {x} → U -perf {x} is fully faithful. Theorem 4.6 and §4.2.4 show
that U -perf {x} is a minimal non-zero thick subcategory of U -perf. The lemma
follows.
Remark 4.22. Bondal and Orlov [BoOr] show that the structure of graded
category of D b (X-coh) (we forget the distinguished triangles) is enough to
reconstruct X in Theorem 4.19.
∼
Given α : X → Y an isomorphism of varieties, we have an equivalence
∼
α∗ : D b (X-coh) → D b (Y -coh). Given L a line bundle on X, we have a self-
∼
equivalence L⊗? : D b (X-coh) → D b (X-coh). Finally, given n ∈ Z, we have
the self-equivalence [n].
This gives a injective morphism from Z × (Pic X Aut(X)) to the group
Aut(D b (X-coh)) of isomorphism classes of self-equivalences of D b (X-coh).
We can now complete Theorem 4.19.
Theorem 4.23 (Bondal-Orlov). Let X be a smooth connected projective variety

−1 ∼
with ωX or ωX ample. Then the canonical map Z × (Pic X Aut(X)) →
Aut(D b (X-coh)) is an isomorphism.
Proof. Let F be a self-equivalence of D b (X-coh). In the proof of Theorem 4.19,

we have constructed an automorphism ψ of X such that F (DZ (X-coh)) =
Dψ(Z) (X-coh) for all closed subsets Z of X. Replacing F by F ψ ∗ , we can
assume that DZ (X-coh) is stable by F for all Z closed in X. Furthermore, F
restricts to a self-equivalence Fx of D b (Ox -mod) for all closed points x of X.
Let C = F (OX ). Then Hom(Cx , Cx [i]) Hom(Ox , Ox [i]) = 0 for i = 0.
Let D be a bounded complex of finitely generated projective Ox -modules
that is quasi-isomorphic to Cx and such that given i minimal (resp. maxi-
mal) with dDi = 0, then dDi is not a split injection (resp. a split surjection).
Let r and s be those minimal and maximal integers. The canonical map
Hom(D r , D s+1 ) → Hom(D, D[s − r]) is non-zero: this is impossible. It fol-
lows that H i (Cx ) is concentrated in a single degree i = rx and H rx (Cx ) is free.
In addition, End(Cx ) End(Ox ) = Ox , hence H rx (Cx ) is free of rank 1. The
set of closed points x with H i (Cx ) = 0 is closed in X. Since X is connected, we
deduce that rx = r is constant. It follows that C Hr (C)[−r] and L = Hr (C)
is a line bundle on X. Replacing F by (L−1 [r] ⊗ −) ◦ F , we can assume that
F (OX ) OX .
We want to prove now that F is isomorphic to the identity functor. By
Orlov’s representability Theorem [Or], there exists K ∈ D b ((X × X)-coh) such
that F Rp∗ (K ⊗L q ∗ (−)), where p, q are the first and second projections
X × X → X. Given x1 , x2 closed points of X, we have Hom(K, O{(x1 ,x2 )} [i])
Hom(F (O{x1 } ), O{x2 } [i]) δx1 ,x2 δ0,i k. It follows that K i∗ M, where M is a
line bundle on X and i : X → X × X is the diagnal embedding. We deduce
that K i∗ OX and F Id.
Remark 4.24. A proof similar to the one of Theorem 4.23 shows that Pic(X)
∼
Aut(X) → Aut(X-coh) for any variety X.
References
[Ba1] P. Balmer, Presheaves of triangulated categories and reconstruction of
schemes, Math. Ann. 324 (2002), 557–580.
[Ba2] P. Balmer, The spectrum of prime ideals in tensor triangulated categories,
J. Reine Angew. Math. 588 (2005), 149–168.
[BaFl] P. Balmer and G. Favi, Gluing techniques in triangular geometry, Q. J. Math.
58 (2007), 415–441.
[BoOr] A. Bondal and D. Orlov, Reconstruction of a variety from the derived category
and groups of autoequivalences, Compositio Math. 125 (2001), 327–344.
[Ca] A. Caldararu, letter to Tom Bridgeland, 5 November 2002.
[Ga] P. Gabriel, Des catgories abéliennes, Bull. Soc. Math. France 90 (1962),
323–448.
[Ho] M. Hopkins, Global methods in homotopy theory, in “Homotopy theory
(Durham, 1985)”, 73–96, London Math. Soc. Lecture Note Ser., 117, Cam-
bridge Univ. Press, 1987.
[KaScha] M. Kashiwara and P. Schapira, “Categories and sheaves”, Springer Verlag,
2006.
[Nee1] A. Neeman, The chromatic tower for D(R), Topology 31 (1992), 519–532.
[Nee2] A. Neeman, The connection between the K-theory localization theorem of
Thomason, Trobaugh and Yao and the smashing subcategories of Bousfield
and Ravenel, Ann. Sci. Ecole Norm. Sup. 25 (1992), 547–566.
[Nee3] A. Neeman, “Triangulated categories”, Princeton University Press, 2001.
[Or] D. Orlov, Equivalences of derived categories and K3 surfaces, J. Math. Sci.
(New York) 84 (1997), 1361–1381.
[Rou1] R. Rouquier, Peccot lectures, Collège de France, March 2000.
[Rou2] R. Rouquier, Dimensions of triangulated categories, Journal of K-theory 1
(2008), 193–256.
[Se] J.-P. Serre, Prolongement des faisceaux analytiques cohérents, Ann. Inst.
Fourier 16 (1966), 363–374.
[Th1] R.B. Thomason, The local to global principle in algebraic K-theory, Pro-
ceedings ICM, Kyoto, 1990, 381–394, Springer Verlag, 1991.
[Th2] R.B. Thomason, The classification of triangulated subcategories, Compositio
Math. 105 (1997), 1–27.
[ThTr] R.B. Thomason and T.F. Trobaugh, Higher algebraic K-theory of schemes
and of derived categories, Grothendieck Festschrift vol. III, 247–435,
Birkhauser, 1990.
[To] B. Toen, The homotopy theory of dg-categories and derived Morita theory,
Invent. Math. 167 (2007), 615–667.
Triangulated categories for the analysts
pi erre s ch api ra
Abstract. This paper aims at showing how the tools of Algebraic Geometry
apply to Analysis. We will review various classical constructions, including Sato’s
hyperfunctions, Fourier-Sato transform and microlocalization, the microlocal
theory of sheaves (with some applications to PDE) and explain the necessity of
Grothendieck topologies to treat algebraically generalized functions with growth
conditions.
Introduction
In this paper, we will show how the tools of Algebraic Geometry– sheaves,
triangulated and derived categories, Grothendieck topologies and stacks– play
(or should play) a crucial role in Analysis.
Note that, conversely, some problems of Analysis led to new algebraic con-
cepts. For example, one of the deepest notion related to triangulated categories
is that of t-structure, and as a particular case, that of perverse sheaves, and
these notions emerged with the study of the Riemann Hilbert correspondence,
a problem dealing with differential equations.
Another example is the Fourier transform, clearly one of the most essential
tools of the analysts, until it was categorified by Sato and applied to algebraic
analysis, next transposed to algebraic geometry (the Fourier-Mukai transform).
The classical analysts are used to work in various functional spaces con-
structed with the machinery of functional analysis and Fourier transform, but
Sato’s construction of hyperfunctions [28] in the 60’s does not use any of these
tools. It is a radically new approach which indeed has entirely modified the
mathematical landscape in this area. The functional spaces are now replaced by
“functorial spaces”, that is, sheaves of generalized holomorphic functions on a
complex manifold X or, more precisely, complexes of sheaves RHom (G, OX ),
where G is an object of the derived category of R-constructible sheaves on the
real underlying manifold to X. Putting general systems of linear partial dif-
ferential equations, i.e., DX -modules, in the machinery, one is led to study
complexes RHom (G, F ), where F = Sol(M) := RHom DX (M, OX ) is the
complex of holomorphic solutions of a coherent DX -module M.
Date: March 5, 2010.

Mathematics Subject Classification: 58G07, 32A45, 32C38, 35A27.
371
372 Pierre Schapira
The main invariant attached to a coherent DX -module M is its characteristic

variety char(M), a closed conic involutive subset of the cotangent bundle T ∗ X.
For a sheaf G on a real manifold there is a similar invariant, the microsupport
SS(G) which describes the directions of non propagation of G, and this set is
again a closed conic involutive subset of the cotangent bundle. In case of F =
Sol(M), it follows from the Cauchy-Kowalevsky theorem that its microsupport
is nothing but char(M). Therefore, in order to study RHom (G, Sol(M)), one
can forget that one is working on a complex manifold X and dealing with DX -
modules, keeping only in mind two geometrical informations, the microsupport
of G and that of F (see [16]).
The study of the characteristic variety of DX -modules naturally leads to
introduce the ring EX of microdifferential operators, a kind of localization of DX
in T ∗ X. The ring EX was first constructed in [29] using Sato’s microlocalization
functor, the Fourier-Sato transform of the specialization functor. We shall briefly
recall here the main steps of these constructions.
Finally, although classical sheaf theory does not allow one to treat usual
spaces of analysis involving growth conditions, these conditions being not of
local nature, we shall show here how it is possible to overcome this difficulty
by using Grothendieck topologies (see [17]).
0.1 Notations
In this paper, we mainly follow the notations of [16].
(i) We denote by k a field and by Db (kX ) the bounded derived category of
sheaves of k-vector spaces on a topological space X. More generally, if A is
a sheaf of rings on X, we denote by Mod(A) the category of left A-modules
and by Db (A) its bounded derived category. If there is no risk of confusion, we
write RHom instead of RHom kX and similarly for RHom and ⊗.
(ii) If Z is a locally closed subset of the topological space X, we denote by
kXZ the sheaf which is the constant sheaf with stalk k on Z and which is 0 on
X \ Z. If there is no risk of confusion, we write kZ instead of kXZ .
(iii) For a real manifold X, we denote by dimX its dimension and by orX the
orientation sheaf. For a morphism of manifolds f : X − → Y , we set orX/Y =
orX ⊗f −1 orY .
(iv) We denote by DX ( • ) = RHom kX ( • , kX ) the duality functor in Db (kX )
and by ωX the dualizing complex. Recall that ωX orX [dimX ].
(v) For a (real or complex) manifold X, we denote by τ : T X − → X and
π : T ∗X − → X its tangent bundle and cotangent bundle, respectively. Let
f:X− → Y be a morphism of (real or complex) manifolds. To f are associated
the maps

TX J
f
/ X ×Y T Y fτ
/ T Y, T ∗ X Lo
fd
X ×Y T ∗ Y
fπ
/ T ∗Y
JJ LLL
JJ LLL π
J
τ JJJ π LLL
τ τ π
J% L%
X
f
/Y X
f
/ Y.
We denote by TX∗ X the zero-section of T ∗ X and by TX∗ Y = fd−1 TX∗ X the conor-
mal bundle to f . If f is an embedding, we identify TX∗ Y to a sub-bundle of
T ∗ Y and call it the conormal bundle to X.
(vi) For a complex manifold X, we denote by OX the sheaf of holomorphic
functions and by #X the sheaf of holomorphic forms of maximal degree. We
denote by dX the complex dimension of X.
1. Generalized functions
In the sixties, people used to work in various spaces of generalized functions
on a real manifold. The situation drastically changed with Sato’s definition of
hyperfunctions [28].
Consider first the case where M is an open subset of the real line R and let
X an open neighborhood of M in the complex line C satisfying X ∩ R = M.
The space B(M) of hyperfunctions on M is given by
B(M) = O(X \ M)/O(X).
It is easily proved that this space depends only on M, not on the choice of X,
and that the correspondence U
→ B(U ) (U open in M) defines a flabby sheaf
BM on M.
Classically, the “boundary value” of a holomorphic function ϕ(z) defined in
the open set X ∩ {Im z > 0} of the complex line, if it exists, is the limit (for a
>
suitable topology) of the function ϕ(x + iy) as y −
→ 0. With Sato’s definition,
the boundary value always exists and is no more a limit. Indeed, it is the
class of the holomorphic function ψ(z) ∈ O(X \ M) given by ψ(z) = ϕ(z) for
Im z > 0 and ψ(z) = 0 for Im z < 0.
On a manifold M of dimension n, the sheaf BM was originally defined as
BM = HM
n
(OX ) ⊗ orM
where X is a complexification of M. Since X is oriented, Poincaré’s dual-

ity gives the isomorphism DX (CM ) orM [−n]. An equivalent definition of
374 Pierre Schapira
hyperfunctions is thus given by
BM = RHom CX (DX (CM ), OX ). (1.1)
The importance of Sato’s definition is twofold: first, it is purely algebraic

(starting with the analytic object OX ), and second it highlights the link between
real and complex geometry.
Let us define the notion of “boundary value” in this settings. Consider a
subanalytic open subset # of X and denote by # its closure. Assume that:

DX (C# ) C# ,
M ⊂ #.
The morphism C# − → CM defines by duality the morphism DX (CM ) − →

DX (C# ) C# . Applying the functor RHom ( • , OX ), we get the boundary
value morphism
b : O(#) −
→ B(M). (1.2)
Sato’s sheaf of hyperfunctions is an example of a sheaf of generalized

holomorphic functions. Another example of such a sheaf is as follows.
Consider a closed complex hypersurface Z of the complex manifold X
and denote by U its complementary. Let j : U %→ X denote the embedding.
Then j∗ j −1 OX represents the sheaf on X of functions holomorphic on U with
possible (essential) singularities on Z. One has
j∗ j −1 OX RHom CX (CU , OX ). (1.3)
Both examples (1.1) and (1.3) are described by a sheaf of the type
RHom (G, OX ), with G a constructible sheaf (see [16] for an exposition).
Recall that a sheaf G on a real analytic manifold X is R-constructible if there
exists a subanalytic stratification of X on which G is locally constant of finite
rank (over the field k). One denotes by DbR−c (kX ) the full triangulated sub-
category of Db (kX ) consisting of objects with R-constructible cohomology.
On a complex manifold, replacing the subanalytic stratifications by complex
analytic stratifications, one also gets the category DbC−c (kX ) of C-constructible
sheaves.
The advantage of considering the category DbR−c (kX ) is that the properties
of being constructible (in the derived sense) is stable by the six Grothendieck
operations (with suitable properness hypotheses).
To summarize, the classical functional spaces are now replaced by the “func-
torial spaces” RHom CX (G, OX ), where G ∈ DbR−c (CX ).
2. D-modules
References for the theory of D-modules are made to [8, 9].
The theory of D-modules appeared in the 70’s with Kashiwara’s thesis [8]
and Bernstein’s paper [2]. However, already in the 60’s, Sato had the main
ideas of the theory in mind and gave talks at Tokyo University on these topics.
Unfortunately, Sato did not write anything and it seems that his ideas were not
understood at this time. (See [1, 30].)
Let X be a complex manifold. One denotes by DX the sheaf of rings of
holomorphic (finite order) differential operators. A system of linear differential
equations on X is a left coherent DX -module. The link with the intuitive notion
of a system of linear differential equations is as follows. Locally on X, M may
be represented as the cokernel of a matrix ·P0 of differential operators acting
on the right. By classical arguments of analytic geometry (Hilbert’s syzygies
theorem), one shows that, locally, M admits a bounded resolution by free
modules of finite type, that is, M is locally isomorphic to the cohomology of
a bounded complex
• Nr N1 N0 ·P0
M := 0 −
→ DX −
→ ··· −
→ DX −→ DX −
→ 0. (2.1)
Let us introduce the notation:
Sol( • ) := RHom DX ( • , OX ). (2.2)
The complex Sol(M) of holomorphic solutions of M may be locally calculated

•
by applying the functor Hom DX ( • , OX ) to M . Hence
N0 N1P0 · Nr
Sol(M) 0 −
→ OX −→ OX −
→ ··· −
→ OX −
→ 0, (2.3)
where now P0 · operates on the left.

One defines naturally the characteristic variety char(M) of a coherent DX -
module M. This is a closed complex analytic subset of T ∗ X, conic with respect
to the action of C× on T ∗ X. For example, if M has a single generator u with
relation Iu = 0, where I is a locally finitely generated ideal of DX , then
char(M) = {(x; ξ ) ∈ T ∗ X; σ (P )(x; ξ ) = 0 for all P ∈ I}
where σ (P ) denotes the principal symbol of P . A fundamental result of

[29] asserts that for a coherent D-module M, char(M) is an involutive (i.e.,
coisotropic) subset of T ∗ X. The proof uses infinite order microdifferential
operators and quantized contact transformations. Of course, the involutivity
theorem has a longer history, including the previous work of Guillemin-Quillen-
Sternberg [6], and culminating with the purely algebraic proof of Gabber [5].
376 Pierre Schapira
We denote by Dbcoh (DX ) the full triangulated subcategory of Db (DX ) consist-

ing of objects with coherent cohomologies. However, we need a refined notion.
A DX -module M is “good” if on each relatively compact open subset U of
X, M is generated by a coherent OX |U -modules. We denote by Dbgood (DX )
the full triangulated subcategory of Db (DX ) consisting of objects with good
cohomologies.
Operations on D-modules
Let f : X −
→ Y be a morphism of complex manifolds. The sheaf
−1
→Y := OX ⊗f −1 OY f DY
DX− (2.4)
is naturally endowed with a structure of a (DX , f −1 DY )-bimodule, but one

shall be aware that the left action of DX is not simply its action on OX (see [9]
for details).
The inverse image functor Df −1 for D-modules is given by
L
Df −1 N = DX− −1
→Y ⊗f −1 DY f N , N ∈ Db (DY ).
One says that f is non characteristic for N ∈ Dbcoh (DY ) if (see Notations 0.1)
fπ−1 char(N ) ∩ (TX∗ Y ) ⊂ X ×Y TY∗ Y.
In this case, Df −1 N ∈ Dbcoh (DX ) and the Cauchy-Kowalevsky-Kashiwara the-

orem asserts that there is a natural isomorphism
∼
f −1 Sol(N ) −→ Sol(Df −1 N ). (2.5)
The proper direct image functor Df ! for (right) D-modules is given by

L op
Df ! M := Rf ! (M⊗DX DX−
→Y ), M ∈ Db (DX ).
One defines the direct image functor Df ! for left D-modules by using the line
bundles #X and #Y (or their inverse) which intertwine the left and right struc-
tures. If M ∈ Dbgood (DX ) and f is proper on the support of M, one deduces from
Grauert’s direct images Theorem that Df ! M belongs to Dbgood (DY ). Moreover,
there is a natural isomorphism
Rf ! Sol(M) [dX ] Sol(Df ! M) [dY ]. (2.6)
(Recall that dX (resp. dY ) is the complex dimension of X (resp. Y ).)

L
The product M⊗OX L of two left DX -modules is naturally endowed with a
D
structure of a left DX -module. We denote it by M⊗L.
Holonomic systems
Since the characteristic variety of a coherent DX -module M is involutive,
it is natural to study with a particular attention the extreme case where this
characteristic variety has minimal dimension.
Definition 2.1. A holonomic system on X is a coherent DX -module whose

characteristic variety is Lagrangian in T ∗ X.
Denote by Dbhol (DX ) the full triangulated subcategory Db (DX ) consist-

ing of objects with holonomic cohomology and recall that we set Sol( • ) =
RHom D ( • , OX ). The first fundamental result of the theory of holonomic D-
modules is Kashiwara’s constructibility theorem:
Theorem 2.2. [10] The functor Sol induces a functor
Sol : (Dbhol (DX ))op −

→ DbC−c (CX ).
Simple examples in dimension one show that this functor is not fully faithful,
but Kashiwara-Oshima [15] and Kashiwara-Kawai [14] gave the definition of
a “regular holonomic” D-module and Kashiwara [11] proved the Riemann-
Hilbert correspondence, that is, the equivalence of categories
∼
(Dbreg−hol (DX ))op −→ DbC−c (CX )
where now Dbreg−hol (DX ) denotes the full triangulated subcategory of Db (DX )
consisting of objects with regular holonomic cohomology.
Integral transforms
In this subsection, we shall study the action of integral transforms on D-modules
and their sheaves of generalized solutions.
Consider two complex manifolds X and Y , of complex dimension dX and
dY respectively, and the correspondence:
X × YG g
f
ww GGG
{www G#
X Y
For G ∈ Db (CY ) and L ∈ Db (CX×Y ), we set
L ◦ G = Rf ! (L ⊗ g −1 G). (2.7)
For M ∈ Db (DX ) and L ∈ Db (DX×Y ), we set

D D
M ◦ L = Dg ! (Df −1 M⊗L). (2.8)
378 Pierre Schapira
Now we consider
M ∈ Dbgood (DX ), L ∈ Dbreg−hol (DX×Y ), G ∈ Db (CY ),
and we make the hypotheses:

−1
f supp(M) ∩ supp(L) is proper over Y ,
(2.9)
(char(M) × TY∗ Y ) ∩ char(L) ⊂ TX×Y
∗
(X × Y ).
Theorem 2.3. (See [4].) Set L = RHom DX×Y (OX×Y , L) and assume (2.9).
One has the isomorphism
D
RHom CY (G, Sol(M ◦ L)) RHom CX (L ◦ G, Sol(M)) [dX − 2dY ].
This result was first stated (in a slightly less general formulation) in [4]. Its
proof makes use of the isomorphisms (2.5) and (2.6).
This result admits several variants: one may replace OX with its temperate
version or formal version (see § 5 below and [17, Ch. 7]), there are twisted
versions and also G-equivariant versions (see [13]).
Many applications of this theorem are exposed in [3], in particular to
projective duality, Penrose transform and, more generally, Grasmanniann
correspondences.
Remark 2.4. The kernel L appearing in Theorem 2.3 is regular holonomic.

However, irregular kernels may naturally appear which makes the situation
much more intricate. The Laplace transform is such an example of irregular
kernel (see [20]).
3. Microsupport
References for this section are made to [16].
Let X denote a real manifold of class C ∞ and let F ∈ Db (kX ). The micro-
support SS(F ) of F is the closed conic subset of T ∗ X defined as follows.
Definition 3.1. Let U be an open subset of T ∗ X. Then U ∩ SS(F ) = ∅ if

and only if for any x0 ∈ X and any real C ∞ -function ϕ : X −
→ R such that
ϕ(x0 ) = 0, dϕ(x0 ) ∈ U , one has:
(Rϕ≥0 (F ))x0 = 0.
In other words, F has no cohomology supported by the closed half spaces

whose conormals do not belong to its microsupport. One again proves that the
microsupport is a closed conic involutive subset of T ∗ X.
Assume now that X is a complex manifold, that we identify with its real
underlying manifold. The link between the microsupport of sheaves and the
characteristic variety of coherent D-modules is given by:
Theorem 3.2. Let M be a coherent D-module. Then SS(Sol(M)) = char(M).
The inclusion SS(Sol(M)) ⊂ char(M) is the most useful in practice. Its

proof only makes use of the Cauchy-Kowalevsky theorem in its precise form
given by Leray (see [23] or [7, § 9.4]) and of purely algebraic arguments.
Now consider a space of generalized functions RHom (G, OX ) associated
with an R-constructible sheaf G, and a system of linear partial differential
equations, that is, a coherent DX -module M. The complex of generalized
functions solution of this system is given by the complex
RHom DX (M, RHom (G, OX )) RHom (G, Sol(M)).
Setting F = Sol(M), we are reduced to study RHom (G, F ). Our only infor-
mation is now purely geometrical, this is the microsupport of G and that of F
(this last one being the characteristic variety of M). Now, we can forget that
we are working on a complex manifold and that we are dealing with partial
differential equations. We are reduced to the microlocal study of sheaves on a
real manifold [16].
Application: ellipticity
Let us show how the classical Petrowsky regularity theorem may be obtained
with the only use of the Cauchy-Kowalevsky-Leray Theorem, and some sheaf
theory.
The regularity theorem for sheaves is as follows.
Theorem 3.3. Let X be a real analytic manifold and let F, G ∈ Db (kX ).

Assume that G is R-constructible and SS(G) ∩ SS(F ) ⊂ TX∗ X. Then the natural
morphism
RHom (G, kX ) ⊗ F −
→ RHom (G, F ) (3.1)
is an isomorphism.
Let us come back to the situation where X is a complexification of a real man-

ifold M and choose k = C. Set G = DX (CM ) and F = RHom DX (M, OX ).
A differential operator P on X is elliptic (with respect to M) if its principal
symbol σ (P ) does not vanish on the conormal bundle TM∗ X, outside of the
zero-section. More generally a coherent DX -module M is elliptic with respect
380 Pierre Schapira
to M if
char(M) ∩ TM∗ X ⊂ TX∗ X.
By Theorem 3.2, SS(F ) ∩ TM∗ X ⊂ TX∗ X. Let AM = CM ⊗ OX denotes the

sheaf of analytic functions on M. The regularity theorem for sheaves gives the
isomorphism
∼
RHom DX (M, AX ) −→ RHom DX (M, BX ).
In other words, the two complexes of real analytic and hyperfunction solutions
of an elliptic system are quasi-isomorphic. This is the Petrowsky’s theorem for
D-modules.
Of course, this result extends to other sheaves of generalized holomorphic
functions, replacing the constant sheaf CM with an R-constructible sheaf G.
For further developments, see [31].
4. Microlocal analysis
For a detailed exposition, see [16].
Fourier-Sato transform
Let τ : E −
→ M be a finite dimensional real vector bundle over a real manifold
M with fiber dimension n, π : E ∗ −→ M the dual vector bundle. Denote by p1
and p2 the first and second projection defined on E ×M E ∗ , and define:
P = {(x, y) ∈ E ×M E ∗ ; x, y ≥ 0}
P = {(x, y) ∈ E ×M E ∗ ; x, y ≤ 0}
Consider the diagram:
E ×M EK∗
p1 tt KKpK2
t K%
y t
t ∗
E KK E
KKK
rrrr
τ K% r
y r π
M
Denote by DbR+ (kE ) the full triangulated subcategory of Db (kE ) consisting of
conic objects, that is, objects F such that for all j , H j (F ) is locally constant
on the orbits of the action of R+ on E. One defines the two functors
( • )∧
DbR+ (kE ) o / Db (k ∗ )
R+ E
( • )∨
by setting for F ∈ DR+ (kE ) and G ∈ DR+ (kE ∗ )
F ∧ = Rp2 ! (p1−1 F )P , G∨ = Rp1 ! (p2! G)P .
The main result of the theory is the following.
Theorem 4.1. The two functors ( • )∧ and ( • )∨ are equivalences of categories

inverse to each other. In particular, for F1 and F2 in DbR+ (kE ), there is a natural
isomorphism:
RHom (F1 , F2 ) RHom (F1∧ , F2∧ ). (4.1)
Example 4.2. (i) For a convex cone γ in E, denote Intγ its interior, by γ a = −γ
its image by the antipodal map and by γ ◦ its polar cone in E ∗ . Then, for a closed
convex proper cone γ with M ⊂ γ ,
(kγ )∧ kIntγ ◦ .
For an open convex cone λ in E:
(kλ )∧ kλ◦a ⊗ orE ∗ /M [−n].
(ii) Let E denote the Euclidian space Rn and let x = (x1 , . . . , xn ) be the coordi-
nates. Let n = p + q with p, q ≥ 1 and set x = (x , x ) with x = (x1 , . . . , xp ),
x = (xp+1 , . . . , xn ). We denote by u = (u , u ) the dual coordinates on the
dual space E ∗ .
Let γ denote the closed solid cone in E,
γ = {x; x − x ≥ 0}
2 2
and let λ denote the closed solid cone in E ∗ :
λ = {u; u − u ≤ 0}.

2 2
We have (see [20]):
k∧γ kλ [−p].
Specialization and microlocalization

Let X be a real manifold (say of class C ∞ ), M a closed submanifold. Denote
→ M and π : TM∗ X −
by τ : TM X − → M the normal bundle and the conormal
bundle to M in X, respectively. Let F ∈ Db (kX ). The specialization of F along
M, denoted νM (F ), is an object of DbR+ (kTM X ). Its cohomology objects are
described as follows. If V is an open cone in TM X, then
H j (V ; νM (F )) lim H j (U ; F )
−→
U
382 Pierre Schapira
where U ranges over the family of open subsets of X which are “tangent” to V ,
that is, open tuboids in X with wedge M whose “profiles” is V . (For a precise
definition, refer to [16, § 4.2].)
The Sato’s microlocalization of F along M, denoted μM (F ), is the Fourier-
Sato transform of νM (F ), an object of DbR+ (kTM∗ X ). It satisfies:
Rπ ∗ μM (F ) RM (F ),
j
H j (μM (F ))(x0 ;ξ0 ) lim HU ∩Z (U ; F ),
−→
U,Z
where, in the last formula, (x0 ; ξ0 ) ∈TM∗ X,

U ranges over the family of open
neighborhoods of x0 in X and Z ranges over the family of closed tuboids in X
with wedge M whose profiles λ in TM X satisfy (x0 ; ξ0 ) ∈ Intλ◦a .
Using the diagonal of X × X, one defines the bifunctor μhom by setting:
μhom (G, F ) = μ RHom (q2−1 G, q1! F )
where qi (i = 1, 2) denotes the i-th projection on X × X. Note that
Rπ ∗ μhom (G, F ) RHom (G, F ),
μhom (kM , F ) μM (F ),
supp μhom (G, F ) ⊂ SS(G) ∩ SS(F ).
Now assume that M is a real analytic manifold and X is a complexification of
M. The sheaf of Sato’s microfunctions on TM∗ X is defined by:
CM = μhom (DX CM , OX ).
(It is proved that this complex is concentrated in degree 0.) Hence, a hyper-
function is nothing but a microfunction globally defined on TM∗ X. Denote by
spec the natural isomorphism:
∼
spec : BM −→ π∗ CM .
If u is an hyperfunction, Sato defines its analytic wave front set as:
WF(u) = supp(spec(u)),
a closed conic subset of TM∗ X. As an application, consider the situation of
the construction of the boundary value morphism in (1.2). Let ϕ ∈ O(#) and
assume that ϕ is solution of a system of differential equations, that is, ϕ ∈
Hom DX (M, OX ) for a coherent DX -module M. Set F0 = Hom DX (M, OX ).
Since the boundary value morphism factorizes through μhom (C# , F0 ) and
SS(F0 ) ⊂ char(M), we get
WF(b(ϕ)) ⊂ TM∗ X ∩ SS(C# ) ∩ char(M). (4.2)
Remark 4.3. A new microlocalization functor μ has been constructed in [21],

taking its values in the category of ind-sheaves on T ∗ X. It is related to the
functor μhom by the formula
μhom (G, F ) RHom (π −1 G, μ(F )).
Microdifferential operators
The sheaf of microfunctions allows us to analyze hyperfunctions microlocally,
that is, in the cotangent bundle. A similar localization may be performed with
respect to the sheaf of differential operators. Let X be a complex manifold of
dimension dX and denote by the diagonal of X × X. The sheaf of microlocal
operators is defined in [29] by
EXR := μ (OX×X
(0,dX )
) [dX ].
(0,dX )
Here, OX×X = OX×X ⊗q −1 OX q2−1 #X . One proves that EXR is concentrated in
2
degree 0 and is naturally endowed with a structure of a sheaf of C-algebras.
Moreover, for G ∈ Db (CX ), the object μhom (G, OX ) is well defined in Db (EXR )
(see [21]). This applies in particular to the sheaf CM of microfunctions.
The algebra EXR is extremely difficult to manipulate, but it contains the C-
algebra EX of microdifferential operators which is filtered and admits a symbol
calculus. When X is affine and U is open in T ∗ X, a section P ∈ ET ∗ X (U ) is
described by its “total symbol”

σtot (P )(x; ξ ) = pj (x; ξ ), m ∈ Z, pj ∈ (U ; OT ∗ X (j )),
−∞<j ≤m
with the condition:

for any compact subset K of U there exists a positive constant
−j
CK such that sup |pj | ≤ CK (−j )! for all j < 0.
K
(Here OT ∗ X (j ) denotes the subsheaf of OT ∗ X of functions homogeneous of

degree j in the fiber variable.) The total symbol of the product is given by the
Leibniz rule. If Q is an operator of total symbol σtot (Q), then
1
σtot (P ◦ Q) = ∂ξα σtot (P )∂xα σtot (Q).
α∈Nn
α!
∼
If ϕ : T ∗ X ⊃ U −→ V ⊂ T ∗ Y is a homogeneous complex symplectic iso-
morphism, it can be locally “quantized” (see [29]), that is, extended as an
isomorphism of algebras
∼
: ϕ∗ ET ∗ X −→ ET ∗ Y .
384 Pierre Schapira
However, this isomorphism is only locally defined and not unique. That is
why, in general, it is not possible to define such sheaves EX of microdifferential
operators on a homogeneous complex symplectic manifold X. However, Kashi-
wara [12] has shown that such a construction was possible when weakening
the notion of a sheaf of algebras by that of an algebroid stack. (Refer to [19]
for an introduction to stacks.)
Further developments and open problem

A complex cotangent bundle T ∗ X is endowed with the sheaf ET ∗ X . This sheaf
is conic for the action of C× on T ∗ X and one can eliminate this homogeneity
by adding an extra central variable . One gets the algebra WT ∗ X over a field
k, a subfield of k = C(()). The algebra WT ∗ X admits a formal version W T ∗ X

over k which is extremely popular and known as a deformation-quantization
algebra. If X is affine, a section P ∈ W T ∗ X (U ) on an open subset U ⊂ T ∗ X is
a series, called its total symbol:

σtot (P )(x; u, ) = pj (x; u)j , m ∈ Z, pj ∈ OT ∗ X (U ), (4.3)
m≤j <∞
and the total symbol of the product P ◦ Q is given by the Leibniz rule:
|α|
σtot (P ◦ Q) = ∂ α (σtot P )∂xα (σtot Q).
n
α! u
α∈N
Note that WT ∗ X has a natural filtration for which has order −1 and W
T ∗ X (0)
may naturally be considered as a deformation of the Poisson algebra OT ∗ X .
By replacing again the notion of a sheaf of algebras by that of an alge-
broid stack, one can define such objects on complex symplectic manifolds
(see [22, 27]).
A natural question would be to perform an analogous construction for
sheaves on real manifolds, that is, to construct a non conic microlocal the-
ory of sheaves on real symplectic manifolds. This problem is closely related to
Mirror Symmetry (see [25]).
5. The use of Grothendieck topologies

References for this section are made to [17, Ch. 7].
Let X be a real analytic manifold. The usual topology on X does not allow
one to treat usual spaces of analysis with the tools of sheaf theory. For example,
the property of being temperate is not local, and there is no sheaf of temperate
distributions. One way to overcome this difficulty is to introduce a Grothendieck
topology on X. Recall that a Grothendieck topology is not a topology, and in

fact is not defined on a space but on a category. The objects of the category
playing the role of the open subsets of the space, it is an axiomatization of
the notion of a covering. A site is a category endowed with a Grothendieck
topology.
We denote by OpX the category whose objects are the open subsets of X and
the morphisms are the inclusions of open subsets. One defines a Grothendieck
topology on OpX by deciding that a family {Ui }i∈I of subobjects of U ∈ OpX
is a covering of U if it is a covering in the usual sense.
Definition 5.1. Denote by OpXsa the full subcategory of OpX consisting of

subanalytic and relatively compact open subsets. The site Xsa is obtained by
deciding that a family {Ui }i∈I of subobjects of U ∈ OpXsa is a covering of U if

there exists a finite subset J ⊂ I such that j ∈J Uj = U .
Let us denote by
ρ: X −
→ Xsa (5.1)
the natural morphism of sites. As usual, we have a pair of adjoint functors

(ρ −1 , ρ∗ ):
Mod(kX ) o
ρ∗
/ Mod(k
Xsa ).
ρ −1
The functor ρ −1 also admits a right adjoint ρ! . For F ∈ Mod(kX ), ρ! F is the

sheaf associated to the presheaf U
→ F (U ), U ∈ OpXsa .
Let us denote by CX∞ the sheaf of complex valued C ∞ -functions on X.
Definition 5.2. Let f ∈ CX∞ (U ). One says that f has polynomial growth at
p ∈ X if it satisfies the following condition. For a local coordinate system
(x1 , . . . , xn ) around p, there exist a sufficiently small compact neighborhood
K of p and a positive integer N such that
N
supx∈K∩U dist(x, K \ U ) |f (x)| < ∞ . (5.2)
It is obvious that f has polynomial growth at any point of U . We say that f is

temperate at p if all its derivatives have polynomial growth at p. We say that
f is temperate if it is temperate at any point.
For an open subanalytic subset U of X, denote by CX∞,t (U ) the subspace of

CX∞ (U ) consisting of temperate functions.
Using Lojasiewicz’s inequalities [24], one easily proves that the presheaf
CX∞,t
sa
, given by U
→ CX∞,t (U ), is a sheaf on Xsa . One calls it the sheaf of
386 Pierre Schapira
temperate C ∞ -functions on Xsa . Note that the sheaf ρ∗ DX does not operate on
CX∞,t
sa
but ρ! DX does.
Now let X be a complex manifold. We still denote by X the real underlying
manifold and we denote by X the complex manifold conjugate to X. One
defines the sheaf of temperate holomorphic functions OX t
sa
as the Dolbeault
∞,t
complex with coefficients in CXsa . More precisely
OX
t
sa
= RHom ρ! DX (ρ! OX , CX∞,t
sa
). (5.3)
Note that the object OX

t
sa
∈ Db (ρ! DX ) is not concentrated in degree zero in
dimension > 1. Nevertheless, it should have many important applications. Let
us mention two of them:
(a) One proves that Sato’s construction of hyperfunctions, when applied to

OXt
sa
, gives the sheaf of Schwartz’s distributions.
(b) Very little is known on irregular holonomic D-modules (see [26]) in
dimension higher than one, and even in dimension one, there are no
(to our opinion) totally satisfactory results. In [18], one calculates the
temperate holomorphic solutions of the DX -module M := DX exp(1/z) =
DX /DX (z2 ∂z + 1), where X = C. The result obtained shows that the sheaf
of temperate holomorphic solutions gives more informations than the clas-
sical sheaf of holomorphic solutions.
The sheaf OXt

sa
is simply an example which shows that the methods of Alge-
braic Analysis may be applied to treat generalized functions with growth
conditions.
References
[1] E. Andronikof, Interview with Mikio Sato, Notices of the AMS, 54 Vol 2, (2007).
[2] I. Bernstein Modules over a ring of differential operators, Funct. Analysis and
Appl. 5 pp 89–101 (1971).
[3] A. D’Agnolo, Sheaves and D-modules in integral geometry, in Contemp. Math.,
251 Amer. Math. Soc., pp 141 (2000).
[4] A. D’Agnolo and P. Schapira, Leray’s quantization of projective duality, Duke
Math. Journ. 84 pp 453–496 (1996).
[5] O. Gabber, The integrability of the characteristic variety, Amer. Journ. Math. 103
pp 445–468 (1981).
[6] V. Guillemin, D. Quillen and S. Sternberg, The integrability of characteristics,
Comm. Pure and Appl. Math. 23 39–77 (1970)
[7] L. Hörmander, The analysis of linear partial differential operators, Grundlehren
der Math. Wiss. 256 Springer-Verlag (1983).
[8] M. Kashiwara, Algebraic study of systems of partial diffential equations, Thesis,

Tokyo Univ. (1970), translated by A. D’Agnolo and J-P. Schneiders, Mémoires
Soc. Math. France 63 (1995).
[9] , D-modules and Microlocal Calculus, Translations of Mathematical
Monographs, 217 American Math. Soc. (2003).
[10] , On the maximally overdetermined systems of linear differential equations
I, Publ. RIMS, Kyoto Univ. 10 pp 563–579 (1975).
[11] , The Riemann-Hilbert problem for holonomic systems, Publ. RIMS, Kyoto
Univ. 20 pp 319–365 (1984).
[12] , Quantization of contact manifolds, Publ. RIMS, Kyoto Univ. 32 p. 1–5
(1996).
[13] , Equivariant derived category and representation of real semisimple Lie
groups. Representation theory and complex analysis, Lecture Notes in Math.
Springer-Verlag, to appear
[14] M. Kashiwara and T. Kawai, On the holonomic systems of microdifferential equa-
tions III, Publ. RIMS, Kyoto Univ. 17 pp 813–979 (1981).
[15] M. Kashiwara and T. Oshima, Systems of differential equations with regular singu-
larities and their boundary value problems, Ann. of Math. 106 (1977), p. 145–200.
[16] M. Kashiwara and P. Schapira, Sheaves on Manifolds, Grundlehren der Math.
Wiss. 292 Springer-Verlag (1990).
[17] , Ind-sheaves, Astérisque 271 Soc. Mathématique de France (2001).
[18] , Microsupport and regularity for ind-sheaves I, Astérique 284 pp 143–164
(2003). ArXiv: math.AG/0108065
[19] , Categories and Sheaves, Grundlehren der Math. Wiss. Springer-Verlag
(2005).
[20] , Integral transforms with exponential kernels and Laplace transform,
Journ. of the A.M.S. 10 pp 939–972 (1997).
[21] M. Kashiwara, P. Schapira, F. Ivorra and I. Waschkies, Microlocalization of ind-
sheaves. in Studies in Lie Theory, Progress in Math. 243 pp 171–221 (2006).
[22] M. Kontsevich, Deformation quantization of algebraic varieties, in: Euro-
Conférence Moshé Flato, Part III (Dijon, 2000) Lett. Math. Phys. 56 (3) p. 271–294
(2001).
[23] J. Leray, Scientific work, Springer-Verlag & Soc. Mathématique de France (1997).
[24] S. Lojaciewicz, Sur le problème de la division, Studia Math. 8 pp. 87–156,
(1961).
[25] D. Nadler and E. Zaslow, Constructible Sheaves and the Fukaya Category,
!"!#$%&
[26] C. Sabbah, Equations différentielles à points singuliers irréguliers et phénomène
de Stokes en dimension 2, Astérisque 263 Soc. Mathématique de France (2000).
[27] P. Polesello and P. Schapira, Stacks of quantization-deformation modules over
complex symplectic manifolds, Int. Math. Res. Notices 49 p. 2637–2664 (2004).
[28] M. Sato, Theory of hyperfunctions, I & II Journ. Fac. Sci. Univ. Tokyo, 8 139–193
487–436 (1959–1960).
[29] M. Sato, T. Kawai, and M. Kashiwara, Microfunctions and pseudo-differential
equations, in Komatsu (ed.), Hyperfunctions and pseudo-differential equations,
Proceedings Katata 1971, Lecture Notes in Math. Springer-Verlag 287 pp. 265–
529 (1973).
388 Pierre Schapira
[30] P. Schapira, Mikio Sato, a visionary of mathematics, Notices of the AMS, 54

Vol 2, (2007).
[31] P. Schapira and J-P. Schneiders, Index theorem for elliptic pairs, Astérisque 224
Soc. Mathématique de France (1994).
Pierre Schapira
Université Pierre et Marie Curie
Institut de Mathématiques
175, rue du Chevaleret, 75013 Paris France
schapira@math.jussieu.fr
http://www.math.jussieu.fr/˜schapira/
Algebraic versus topological
s tefan s ch w ede
These are extended and updated notes of a talk, the first version of which I gave
at the Workshop on Triangulated Categories at the University of Leeds, August
13–19, 2006. These notes are mostly expository and do not contain all proofs;
I intend to publish the remaining details elsewhere.
The most commonly known triangulated categories arise from chain com-
plexes in an abelian category by passing to chain homotopy classes or inverting
quasi-isomorphisms. Such examples are called ‘algebraic’ because they orig-
inate from abelian (or at least additive) categories. Stable homotopy theory
produces examples of triangulated categories by quite different means, and in
this context the source categories are usually very ‘non-additive’ before pass-
ing to homotopy classes of morphisms. Because of their origin I refer to these
examples as ‘topological triangulated categories’.
In this note I want to explain some systematic differences between these
two kinds of triangulated categories. There are certain properties – defined
entirely in terms of the triangulated structure – which hold in all algebraic
examples, but which fail in some topological ones. These differences are all
torsion phenomena, and rationally there is no difference between algebraic and
topological triangulated categories.
A triangulated category is algebraic in the sense of Keller [Ke, 3.6] if it is
triangle equivalent to the stable category of a Frobenius category, i.e., an exact
category with enough injectives and enough projectives in which injectives
and projectives coincide. Examples include all triangulated categories which
one should reasonably think of as ‘algebraic’: various homotopy categories
and derived categories of rings, schemes and abelian categories; stable module
categories of Frobenius rings; derived categories of modules over differential
graded algebras and differential graded categories. By a theorem of Porta [Po,
Thm. 1.2], every algebraic triangulated category which is well generated (a
mild restriction on its ‘size’, see [Ne, Def. 8.1.6 and 8.1.7]) is equivalent
Date: 2000 AMS Math. Subj. Class.: 18E30, 55P42.
389
390 Stefan Schwede
to a localization of the derived category D(A) of a small differential graded

category A.
For an object X of a triangulated category T and a natural number n we
write n · IdX or simply n · X for the n-fold multiple of the identity morphism
in the group [X, X] of endomorphisms in T . We let X/n denote any cone of
n · IdX , i.e., any object which is part of a distinguished triangle
n·
X −→ X −→ X/n −→ X[1] .
A short diagram chase shows that the group [X/n, X/n] is always annihilated
by n2 ; in algebraic triangulated categories, more is true:
Proposition 1. If T is algebraic, then n · X/n = 0.
Proof. We exploit that algebraic triangulated categories are tensored over

Db (Z), the bounded derived category of finitely generated abelian groups.
This means that there is a biexact functor
⊗L : Db (Z) × T −→ T
which is associative and unital up to coherent isomorphism with respect to the

derived tensor product in Db (Z). In Db (Z) we have a distinguished triangle
n·
Z −→ Z −→ Z/n −→ Z[1]
which becomes a distinguished triangle in T after tensoring with X. So Z/n ⊗L

X is isomorphic to X/n. Since the tensor product is additive in each variable
and since n · Z/n = 0 in Db (Z), we conclude that n · (Z/n ⊗L X) = 0.
In contrast to Proposition 1, in a general triangulated category we can have

n · X/n = 0 for suitable choices of X and n. An example arises in the Spanier-
Whitehead category which we now review. The Spanier-Whitehead category
is made from homotopy classes of continuous maps between certain kinds of
topological spaces, and it is the prime example of a topological triangulated
category (to be defined below) which is not algebraic. The Spanier-Whitehead
category was originally introduced (without the formal desuspensions) in [SW].
We recall that the reduced suspension X of a space X with basepoint x0
is given by
X × [0, 1]
X = .
X × {0, 1} ∪ {x0 } × [0, 1]
For example we have S n−1 = ∼ S n , i.e., the suspension of a sphere is homeo-

morphic to a sphere of the next dimension.
Definition 2. The Spanier-Whitehead category, denoted SW, has as objects
the pairs (X, n) where X is a finite CW-complex equipped with a distinguished
basepoint and n ∈ Z. Morphisms are given by
SW((X, n), (Y, m)) = colimk→∞ [ k+n X, k+m Y ]
where square brackets [−, −] denote pointed homotopy classes of continuous,
basepoint preserving maps. The colimit is formed by iterated suspensions;
by Freudenthal’s suspension theorem, it is actually attained at a finite stage.
Composition in SW is defined by composition of representatives, suitably
suspended so that composition is possible.
It is convenient to identify a finite pointed CW-complex X with the object

(X, 0) of the Spanier-Whitehead category. Then two CW-complexes become
isomorphic in SW if and only if they become homotopy equivalent after a
finite number of suspensions.
The Spanier-Whitehead category is triangulated in the following way:
r The shift functor is given by (X, n)[1] = (X, n + 1). Tautologically, the
identity map of n+1 X is an isomorphism between (X, n) and (X, n + 1)
in SW, so suspension is in fact isomorphic to the shift and is invertible in
SW.
r The Spanier-Whitehead category is additive: for pointed spaces X and Y , the
set [X, Y ] of pointed homotopy classes from a suspension has a natural
group structure as follows. The product of f, g : X −→ Y is represented
by the composite
here f
here g
f ∨g
X
pinch
/ X ∨ X / Y
392 Stefan Schwede
For X = S 0 we have X = ∼ S 1 and this reduces to the group structure on the

fundamental group [S 1 , Y ] = π1 (Y, y0 ). On a double suspension as source
object, this group structure is abelian; an example of this is that for n ≥ 2
the higher homotopy groups πn (Y, y0 ) = [S n , Y ] are abelian. In the Spanier-
Whitehead category, every object is a double suspension, so the homomor-
phism sets in SW are naturally abelian groups.
r Mapping cone sequences give distinguished triangles: the mapping cone of
a pointed map f : X −→ Y is the space
X × [0, 1] ∪X×{1} Y
C(f ) =
X × {0} ∪ {x0 } × [0, 1]
CX Y
There is an inclusion i : Y −→ C(f ) and a projection p : C(f ) −→ X

(which collapses Y to a point). In the sequence of pointed maps
f i p
X −→ Y −
→ C(f ) −→ X
the composite of any two is null-homotopic, and these sequences and their
(de-)suspensions give the distinguished triangles in SW.
A verification of the axioms of a triangulated category (except for the octahe-

dral axiom), and more details on the Spanier-Whitehead category can be found
in Ch. 1, §2 of Margolis book [Mar]. (Margolis does not impose any finiteness
restriction on the objects of the Spanier-Whitehead category; the category SW
is denoted SWf in [Mar].)
The group
SW(S n , X) = colimk→∞ [S k+n , k X] = colimk πk+n ( k X, ∗)
is denoted by πns X and called the n-th stable homotopy group of X. For X = S 0 ,
we abbreviate πns S 0 to πns and speak about the n-th stable homotopy group of
spheres, also called the n-th stable stem. For example, π0s = Z, generated
by the identity map, and π1s = Z/2 generated by the class of the Hopf map
η : S 3 −→ S 2 . The stable stems are easy to define, but notoriously hard to
compute; for example, there is no finite CW-complex X which is non-trivial
in SW for which all stable homotopy groups are known! Much machinery of
algebraic topology has been developed to calculate such groups and understand
their structure, but no one expects to ever get explicit formulae for all stable
homotopy groups of spheres.
We also need the symmetric monoidal smash product on the Spanier-

Whitehead category. This arises from the geometric smash product X ∧ Y
of two pointed spaces X and Y defined as X ∧ Y = (X × Y )/(X × {y0 } ∪
{x0 } × Y ). Suspension is an example of the smash product, i.e., X is nat-
urally homeomorphic to X ∧ S 1 . The smash product can be extended to the
Spanier-Whitehead category by (X, n) ∧ (Y, m) = (X ∧ Y, n + m), and then
it becomes biexact, i.e., an exact functor of triangulated categories in each
variable.
We denote by S = (S 0 , 0) the unit object of the smash product in SW
and refer to it as the sphere spectrum. We use this terminology because the
Spanier-Whitehead category can be identified with the compact objects in a
larger triangulated category, the stable homotopy category, whose objects are
called spectra. However, the differences between algebraic and topological
triangulated categories already show up in the Spanier-Whitehead category,
which is easier to define than the full stable homotopy category.
For n ≥ 2, the mod-n Moore spectrum is defined as a cone of multiplication
by n on the sphere spectrum, i.e., it is part of a distinguished triangle
n·
S −→ S −→ S/n −→ S[1] . (3)
More concretely we can define S/n = (S 1 ∪n D 2 , −1), the formal desuspen-
sion of a two-dimensional mod-n Moore space, i.e., the space obtained from the
circle S 1 by attaching a 2-disc along the degree n map S 1 −→ S 1 , z
→ zn . For
example, S/2 = (RP 2 , −1), the formal desuspension of 2-dimensional real
projective space. If n and m are coprime, then S/(nm) is isomorphic in the
Spanier-Whitehead category to the smash product S/n ∧ S/m. So we often
concentrate on Moore spectra for primes or prime powers, since the other
Moore spectra are smash products of these.
Proposition 1 and the following proposition show that the Spanier-
Whitehead category is not algebraic.
Proposition 4. The morphism 2 · S/2 is nonzero in SW.
The standard tool for proving Proposition 4 and its generalizations below
are cohomology operations, which we quickly review. The mod-p cohomology
of an object (X, n) in SW defined by
H k ((X, n), Fp ) = H̃ k−n (X, Fp )
where the right hand side is reduced singular cohomology with coefficients in
the field Fp . These cohomology groups have a natural action (essentially by
definition) by the mod-p Steenrod algebra, i.e., the algebra of stable, natural,
graded mod-p cohomology operations. The Steenrod algebra is generated for
394 Stefan Schwede
p = 2 by operations Sq i of degree i for i ≥ 1 and for odd p by the Bockstein

operation β of degree 1 and operations P i of degree i(2p − 2) for i ≥ 1. For
p = 2, the operation Sq 1 equals the Bockstein operation. These operations
satisfy the Adem relations, which we do not reproduce here.
The mod-n Moore spectrum is characterized up to isomorphism in the
Spanier-Whitehead category by the property that its integral spectrum homol-
ogy is concentrated in dimension zero where it is isomorphic to Z/n. For a
prime p the mod-p cohomology of S/p is one-dimensional in dimensions 0
and 1, and trivial otherwise, and the Bockstein operation is non-trivial from
dimension 0 to dimension 1.
Proof of Proposition 4. This proposition is a classical fact, which should prob-

ably be credited to Steenrod since the standard proof uses mod-2 Steenrod
operations. We argue by contradiction and suppose that 2 · S/2 = 0. If we
smash the defining triangle (3) with another copy of the mod-2 Moore spec-
trum and use that S is the unit of the smash product, we obtain a distinguished
triangle
2·
S/2 −→ S/2 −→ S/2 ∧ S/2 −→ S/2[1] .
Under our assumption the first map is trivial so the smash product S/2 ∧ S/2
splits in SW as a sum of S/2 and S/2[1]. Thus as a module over the Steenrod
algebra, the mod-2 cohomology of the smash product S/2 ∧ S/2 decomposes
into a sum of two non-trivial summands.
On the other hand, there is a Künneth isomorphism for the mod-2 cohomol-
ogy of a smash product

H n (X ∧ Y, F2 ) ∼
= H p (X, F2 ) ⊗ H q (Y, F2 ) .
p+q=n
This is an isomorphism as modules over the Steenrod algebra with action on

the right hand side given by the Cartan formula

m
Sq m (x ⊗ y) = Sq i x ⊗ Sq m−i y .
i=0
In the mod-2 cohomology of the mod-2 Moore spectrum the Bockstein oper-
ation Sq 1 : H 0 (S/2, F2 ) −→ H 1 (S/2, F2 ) is non-zero. The Cartan formula
shows that the operation Sq 2 is then non-trivial on the tensor product of
two copies of the generator of H 0 (S/2, F2 ). Thus the mod-2 cohomology
of S/2 ∧ S/2 is a 4-dimensional, indecomposable module over the Steenrod
algebra. We have reached a contradiction, which means that we must have
2 · S/2 = 0 in SW.
In topological triangulated categories, the phenomenon that we can have n ·

X/n = 0 is entirely 2-local. To explain this, I have to be more precise about what
I mean by a topological triangulated category. A model category (in the sense of
Quillen [Q]) is an axiomatic framework for homotopy theoretic constructions.
Among other things, a model category structure allows one to define mapping
cones and suspensions and talk about homotopies between morphisms. A model
category is called stable if it has a zero object and the suspension functor is a
self-equivalence of its homotopy category. The homotopy category of a stable
model category is naturally a triangulated category (cf. [Ho1, 7.1.6]); the proof
is essentially the same as for the Spanier-Whitehead category. By definition the
suspension functor is a self-equivalence, and it defines the shift functor. Since
every object is a two-fold suspension, hence an abelian co-group object, the
homotopy category of a stable model category is additive. The distinguished
triangles are defined by mapping cone sequences.
For us a topological triangulated category is any triangulated category which
is equivalent to a full triangulated subcategory of the homotopy category of a
stable model category. An important example which was already mentioned
above is the stable homotopy category of algebraic topology which was first
introduced by Boardman (unpublished; accounts of Boardman’s construction
appear in [Vo] and [Ad, Part III]). There is an abundance of models for the
stable homotopy category, see for example [BF, EKMM, HSS, MMSS, Ly].
The Spanier-Whitehead category SW is equivalent to the full subcategory
of compact objects in the stable homotopy category, so it is a topological
triangulated category in our sense. Further examples of topological triangulated
categories are ‘derived’ (i.e., homotopy) categories of structured ring spectra,
equivariant and motivic stable homotopy categories, sheaves of spectra on a
Grothendieck site or (Bousfield-) localizations of all these, see [SS, Sec. 2.3]
for more details. The theorem of Porta mentioned above has an analogue in
this context: Heider essentially shows in [He, Thm. 4.7] that every topological
triangulated category which is well generated is equivalent to a localization of
the homotopy category Ho(R-mod) of a small spectral category R.
Algebraic triangulated categories are typically also topological (the con-
verse is not generally true, and that is the point of these notes). For algebraic
triangulated categories which are derived categories of abelian categories, this
follows whenever there is a model structure on the category of chain com-
plexes with quasi-isomorphisms as weak equivalences (see for example [Ho2]
or Section 2.4 of [SS] for more details and references). Similarly, for modules
over a Frobenius ring, there is a stable model structure with stable equivalences
as the weak equivalences, see [Ho1, Thm. 2.2.12]. More generally, any well-
generated algebraic triangulated category is equivalent to a localization of the
396 Stefan Schwede
derived category D(A) of small differential graded category A [Po, Thm. 1.2].
The localization can be realized as a Bousfield localization of the ordinary
(i.e., ‘projective’) model structure on modules over A; hence D(A) and its
localizations are topological.
Examples of triangulated categories which are neither algebraic nor topo-
logical were recently constructed by Muro, Strickland and the author [MSS].
The simplest one is the category F(Z/4) of finitely generated free modules
over the ring Z/4. The category F(Z/4) has a unique triangulation with the
identity shift functor and such that the triangle
2 2 2
Z/4 −→ Z/4 −→ Z/4 −→ Z/4
is exact. For the proof of this and an argument why F(Z/4) is not topological
I refer to [MSS]. At present, I do not know any ‘exotic’ (i.e., non-topological
and non-algebraic) triangulated category in which 2 is invertible.
Unlike algebraic triangulated categories, we can not usually expect that a
topological triangulated category can be tensored over Db (Z), the bounded
derived category of finitely generated abelian groups. The appropriate replace-
ment for Db (Z) is the Spanier-Whitehead category: for every topological trian-
gulated category T , there is a biexact pairing
∧ : SW × T −→ T
which is associative and unital up to coherent natural isomorphism with respect

to the smash product in the Spanier-Whitehead category. The shift functor in
T is isomorphic to smashing with the circle S 1 , view as an object of SW.
We use the above pairing as a black box, but I’ll briefly indicate how it can
be constructed. One way is to start from the action of the homotopy category of
pointed simplicial sets on the homotopy category of a pointed model category
C,
∧ : HoSSet∗ × Ho C −→ Ho C ,
which uses a technique called ‘framings’, see [Ho1, Thm. 5.7.3]. The geometric
realization of a finite simplicial set is a finite CW-complex, and up to homo-
topy equivalence, every finite CW-complex arises in this way. Since moreover
suspension is invertible in Ho C, this extends to a well-defined pairing on the
Spanier-Whitehead category.
There is an alternative construction for spectral model categories, i.e., model
categories C which are enriched over the category Sp of symmetric spectra
of [HSS], compatibly with the smash product and the stable model structure
(see [SS, Def. 3.5.1] for details). The derived smash product between symmetric
spectra and C descends to an associative and unital pairing

∧ : Ho(Sp ) × Ho C −→ Ho C ,
which can be restricted to the full subcategory SW of compact objects in
Ho(Sp ). A general stable model category is Quillen-equivalent to a spectral
model category, under mild technical hypothesis (see [SS, Thm. 3.8.2], [Du,
Prop. 5.5 (a) and 5.6 (a)] and [Ho3, Thm. 9.1 and 8.11] for different sets of
sufficient conditions).
In algebraic examples the action of the Spanier-Whitehead category SW
and the bounded derived category Db (Z) (which we used in the proof of
Proposition 1) are related as follows. The chain functor C∗ : SW −→ D b (Z)
associates to an object (X, n) of the Spanier-Whitehead category the reduced
singular chain complex of the CW-complex X, shifted up n dimensions. This
chain functor is strong symmetric monoidal, i.e., there are associative, unital
and commutative isomorphisms C∗ (X ∧ Y ) ∼ = C∗ (X) ⊗L C∗ (Y ) in D b (Z). If
T is an algebraic triangulated category which is also topological, then the
composite
C∗ ×Id ⊗L
SW × T −−−→ Db (Z) × T −−→ T
is naturally isomorphic to the smash product pairing.
Now we can make precise in which sense the possibility of having n · X/n =
0 in topological triangulated categories is a 2-local phenomenon.
Proposition 5. If T is a topological triangulated category and p an odd prime,
then p · X/p = 0 for every object X of T .
Proof. The key point is that if n ≡ 2 mod 4, then the mod-n Moore spectrum
S/n is annihilated by n (which is not the case for n = 2). We recall the easy
proof for odd n, which includes all odd primes. If we take homotopy groups of
the defining triangle (3) (i.e., apply [S, −], where [−, −] denotes morphisms
in SW) we obtain an exact sequence
n· n·
π1s −→ π1s −→ π1 (S/n) −→ π0s −→ π0s −→ π0 (S/n) −→ 0 ,
using that the stable homotopy groups of spheres vanish in negative dimensions.
Since π0s ∼
= Z and π1s ∼ = Z/2 we deduce that π0 (S/n) is cyclic of order n and
π1 (S/n) = 0 for odd n. Now we apply [−, S/n] to the triangle (3) and we
s
obtain a short exact sequence

0 −→ [S[1], S/n] ⊗ Z/n −→ [S/n, S/n] −→ n [S, S/n] −→ 0
where n A = {a ∈ A | na = 0} denotes the group of n-torsion points in an
abelian group A. By the above calculations, the group [S[1], S/n] ⊗ Z/n
398 Stefan Schwede
vanishes and so the group [S/n, S/n] is cyclic of order n for n odd, hence
n · S/n = 0. (For n = 2 we have π1 (S/2) ∼= Z/2 and the analogous short exact
sequence does not split by Proposition 4. So [S/2, S/2] ∼
= Z/4.)
The rest of the argument is then the same as in Proposition 1. For an odd
prime p we can smash the distinguished triangle
p·
S −→ S −→ S/p −→ S[1]
in SW with the object X and obtain a distinguished triangle in T which

shows that S/p ∧ X is isomorphic to X/p. Since p · S/p = 0 in SW and ∧ is
biadditive, we conclude that p · X/p = p · (S/p ∧ X) = 0.
We have seen in Proposition 4 that p · X/p can be non-zero for p = 2. On

the other hand, all triangulated categories that I know have the property that
p · X/p = 0 for odd primes p and all objects X. This leaves us with the
Open problem 6. Let p be an odd prime. Find a triangulated category T and

an object X of T such that p · X/p = 0, or prove that in every triangulated
category T we always have p · X/p = 0.
From what we have discussed so far, it is still conceivable that every topo-
logical triangulated category in which 2 is invertible is algebraic. In particular
one can wonder whether the Spanier-Whitehead category is algebraic after
localization at an odd prime. We now describe a property of triangulated cate-
gories which distinguishes topological from algebraic examples away from the
prime 2.
As before we denote by K/n any cone of n · IdK , which comes as part of a
distinguished triangle
n· π
K −
→ K −→ K/n −→ K[1] .
An extension of a morphism f : K −→ Y is then a morphism f¯ : K/n −→ Y

satisfying f¯π = f . Such an extension exists if and only if n · f = 0, and then
the extension will usually not be unique.
Proposition 7. Let T be an algebraic triangulated category, X an object of

T and n ≥ 2. Then every morphism f : K −→ X/n has an extension f¯ :
K/n −→ X/n such that some (hence any) mapping cone of f¯ is annihilated
by n.
Proof. As in Proposition 1, a choice of model for T as the stable category

of a Frobenius category gives a biexact, associative and unital pairing ⊗L :
Db (Z) × T −→ T and we can take X/n = Z/n ⊗L X and K/n = Z/n ⊗L K.
We define the extension f¯ as the composite

Z/n⊗f μ⊗X
Z/n ⊗L K −−−−→ Z/n ⊗L Z/n ⊗L X −−→ Z/n ⊗L X
where μ : Z/n ⊗L Z/n −→ Z/n is the multiplication map which makes Z/n
into a ring. We choose a distinguished triangle
f¯ ϕ δ
Z/n ⊗L K −→ Z/n ⊗L X −→ C(f¯) −
→ Z/n ⊗L K[1] (8)
and show that the mapping cone C(f¯) of f¯ is annihilated by n.

We consider the diagram
Z/n ⊗L Z/n ⊗L K
Z/n⊗f¯
/ Z/n ⊗L Z/n ⊗L X Z/n⊗ϕ
/ Z/n ⊗L C(f¯) Z/n⊗δ
/ Z/n ⊗L Z/n ⊗L K[1]
μ⊗K μ⊗X σ μ⊗K[1]

Z/n ⊗L K / Z/n ⊗L X / C(f¯) / Z/n ⊗L K[1]
f¯ ϕ δ
whose lower row is the distinguished triangle (8) and whose upper row is (8)
tensored from the left with Z/n. The left square commutes since the multipli-
cation morphism μ is associative in Db (Z). Since both rows are distinguished
triangles, there exists a morphism σ : Z/n ⊗L C(f¯) −→ C(f¯) making the
middle and right square commute.
We consider the morphism
π ⊗ C(f¯) : C(f¯) ∼
= Z ⊗L C(f¯) −→ Z/n ⊗L C(f¯)
which satisfies the two relations
σ (π ⊗ C(f¯))ϕ = σ (Z/n⊗ ϕ)(π ⊗ Z/n ⊗ X) = ϕ(μ⊗ X)(π ⊗ Z/n ⊗X) = ϕ

(9)
and
δσ (π ⊗ C(f¯)) = (μ ⊗ K[1])(Z/n ⊗ δ)(π ⊗ C(f¯)) (10)

= (μ ⊗ K[1])(π ⊗ Z/n ⊗ K[1])δ = δ .
We claim that the morphism
σ = 2σ − σ (π ⊗ C(f¯))σ : Z/n ⊗L C(f¯) −→ C(f¯)
is a retraction to π ⊗ C(f¯). Indeed, by (9) the composite of ϕ : Z/n ⊗L X −→

C(f¯) with the morphism σ (π ⊗ C(f¯)) − Id : C(f¯) −→ C(f¯) becomes triv-
ial, so there exists a morphism g : Z/n ⊗L K[1] −→ C(f¯) such that
400 Stefan Schwede
gδ = σ (π ⊗ C(f¯)) − Id. But then we have
σ (π ⊗ C(f¯)) = 2σ (π ⊗ C(f¯)) − (Id +gδ)σ (π ⊗ C(f¯))

= σ (π ⊗ C(f¯)) − gδσ (π ⊗ C(f¯)) = σ (π ⊗ C(f¯)) − gδ = Id ,
as claimed, where the third equality uses (10).

Now we have shown that C(f¯) is a direct summand of Z/n ⊗L C(f¯), which
is annihilated by n, and so we have n · C(f¯) = 0.
In Theorem 16 below we prove a generalization of Proposition 7, by a

different method.
Here is an example showing that in the situation of Proposition 7 in general
there may not exist any extension f¯ : K/n −→ X/n of f whose mapping cone
is annihilated by n. Proposition 7 and the following proposition show that the
Spanier-Whitehead category localized at 3 is not algebraic.
s ∼
We let β1 ∈ π10 = Z/6 be an element of order 3 (so β1 generates the 3-
primary component of the 10-dimensional stable stem). One way to define β1 is
as the unique element of the Toda bracket α1 , α1 , α1 where α1 ∈ π3s ∼
= Z/24
is the 3-primary part of the element represented by the second Hopf map
ν : S 7 −→ S 4 . We let β̃1 : S[11] −→ S/3 be any lift in the Spanier-Whitehead
category SW of β1 to the mod-3 Moore spectrum S/3. So β̃1 is a morphism
whose composite with the connecting map S/3 −→ S[1] equals the shift of β1 .
We will need below that any such lift β̃1 is detected by P 3 in the sense that in
any mapping cone of β̃1 the Steenrod operation P 3 is a non-trivial isomorphism
from mod-3 cohomology in dimension 0 to dimension 12 (see page 60 of [To,
§5]).
Proposition 11. There is no extension of β̃1 to a morphism β̄ : S/3[11] −→

S/3 whose mapping cone is annihilated by 3.
Proof. We argue by contradiction and suppose that there exists an extension
β̄ : S/3[11] −→ S/3
of β̃1 and a distinguished triangle

β̄ δ
S/3[11] −
→ S/3 −→ C(β̄) −
→ S/3[12]
with 3 · C(β̄) = 0.
Since the stable stems in dimension 21 and 22 consist only of torsion
prime to 3 we have π22 (S/3) = [S[22], S/3] = 0. So there is a morphism
a : S[23] −→ C(β̄) lifting β̃1 in the sense that δa = β̃1 [12]. Since we assumed
3 · C(β̄) = 0, the morphism a can be extended to a morphism ā : S/3[23] −→
C(β̄). We let C(ā) be a mapping cone of ā arising as part of a distinguished

triangle
ā δ
S/3[23] −→ C(β̄) −→ C(ā ) −
→ S/3[24] .
Since the stable stems in dimension 21, 22, 33 and 34 consist only of torsion
prime to 3 we have π34 C(β̄) = [S[34], C(β̄)] = 0 and so there is a morphism
b : S[35] −→ C(ā) lifting β̃1 in the sense that δb = β̃1 [24].
Now we bring cohomology operations into the game to reach a contra-
diction. The Moore spectrum S/3 has its mod-3 cohomology concentrated in
dimensions 0 and 1, where it is 1-dimensional. The mapping cone of b is built
by distinguished triangles from three shifted copies of S/3 and one shifted copy
of S, so its mod-3 cohomology is concentrated in dimensions 0, 1, 12, 13, 24,
25 and 36, where it is 1-dimensional.
The morphism β̃1 is detected by the Steenrod operation P 3 , i.e., in the mod-3
cohomology of the mapping cone of β̃1 the operation P 3 is non-trivial from
dimension 0 to dimension 12. The stable ‘cells’ (i.e., shifted copies of S) of
the mapping cone of b in dimension 12, 24 and 36, are attached to the two
cells directly below by β̃1 , so in the cohomology of the mapping cone of b,
the 3-fold iterate of P 3 is non-trivial from dimension 0 to 36. By the Adem
relations we have (P 3 )3 = (P 7 P 1 − P 8 )P 1 . Since the cohomology is trivial
in dimension 4 the operation P 1 acts trivially, hence so does (P 3 )3 . We have
obtained a contradiction, and so no extension of β̃1 has a mapping cone which
is annihilated by 3.
In topological triangulated categories, the phenomenon that a morphism

f : K −→ X/n may not have any extension whose cone is annihilated by n is
entirely 2- and 3-local, in the following sense.
Proposition 12. If T is a topological triangulated category and p is a prime

bigger than 3, then every morphism f : K −→ X/p has an extension f¯ :
K/p −→ X/p whose mapping cone is annihilated by p.
Proof. The key point is that if n is prime to 6, then the mod-n Moore spectrum
has an associative multiplication in the Spanier-Whitehead category (the mul-
tiplication is also commutative, but that is not relevant for the current proof).
Again I cannot refrain from giving the simple proof below. In contrast, the mod-
2 Moore spectrum does not have a multiplication; the mod-3 Moore spectrum
has a commutative multiplication, but that is not associative in SW, see [To].
As in the proof of Proposition 5, the condition that n is odd guarantees that
n · S/n = 0. So there exists a morphism μ : S/n ∧ S/n −→ S/n which splits
402 Stefan Schwede
the two ‘inclusions’. We show that μ is associative. In fact, the associator
μ(μ ∧ Id − Id ∧μ) : S/n ∧ S/n ∧ S/n −→ S/n
factors as a composite
δ∧δ∧δ
S/n ∧ S/n ∧ S/n −−−→ S[3] −→ S/n
where δ : S/n −→ S[1] is the connecting morphism. We have π3s ∼ = Z/24 and
π2s ∼
= Z/2, so if n is prime to 6 we have [S[3], S/n] = π3 (S/n) = 0. Thus the
associator is trivial or, equivalently, μ is associative in SW.
The rest of the argument is now essentially the same as in Proposition 7. We
can smash the distinguished triangle
n· δ
S −→ S −→ S/n −
→ S[1]
in SW with the object X and obtain a distinguished triangle in T which

shows that S/n ∧ X is isomorphic to X/n. We use the multiplication μ :
S/n ∧ S/n −→ S/n to define the extension f¯ as the composite
Id ∧f μ∧Id
S/n ∧ K −−−→ S/n ∧ S/n ∧ X −−→ S/n ∧ X .
Then we use the same reasoning as in the proof of Proposition 7 to obtain the
retraction σ : S/n ∧ C(f¯) −→ C(f¯) which shows that n · C(f¯) = 0.
This raises the following question:
Open problem 13. Consider a prime p ≥ 5. Does there exist a triangulated

category T , an object X of T and a morphism f : K −→ X/p which does not
admit any extension f¯ to K/p whose mapping cone is annihilated by p?
Remark 14. The proof of Proposition 12 shows that the special features of topo-
logical over algebraic triangulated categories are closely related to existence
and properties of multiplications on mod-n Moore spectra. For primes p ≥ 5,
the mod-p Moore spectrum has a multiplication in the Spanier-Whitehead
category which is commutative and associative. So on the level of tensor trian-
gulated categories, there does not seem to be any qualitative difference between
the Moore spectrum S/p as an object of the Spanier-Whitehead category SW
and Z/p as an object of Db (Z), as long as p ≥ 5. However, mod-n Moore
spectra are never A∞ ring spectra, but rigorously defining what that means
and proving it would lead us too far afield. Theorem 17 below explains how
the higher order non-associativity eventually manifests itself in the triangulated
structure of the Spanier-Whitehead category (i.e., without any reference to the
smash product).
From what we have discussed so far, it is still conceivable that for primes
p ≥ 5 the p-local Spanier-Whitehead category is algebraic. We will now intro-
duce an invariant which we then use to show that this is not the case for any
prime p.
Definition 15. Consider a triangulated category T and a natural number n ≥ 2.

We define the n-order for objects Y of T inductively.
r Every object has n-order greater or equal to 0.

r For k ≥ 1, an object Y has n-order greater or equal to k if and only if for every
object K of T and every morphism f : K −→ Y there exists an extension
f¯ : K/n −→ Y such that some (hence any) mapping cone of f¯ has n-order
greater or equal to k − 1.
We write n -ord(Y ), or n -ordT (Y ) if we need to specify the ambient trian-

gulated category, for the n-order of Y , i.e., the largest k (possibly infinite) such
that Y has n-order greater or equal to k. We define the n-order of the triangu-
lated category T as the n-order of some (hence any) zero object, and denote
it by n -ord(T ). We make some observations which are direct consequences of
the definitions.
r The n-order for objects is invariant under isomorphism and shift.

r An object Y has positive n-order if and only if every morphism f : K −→ Y
has an extension to K/n, which is equivalent to n · f = 0. So n -ord(Y ) ≥ 1
is equivalent to the condition n · Y = 0.
r The n-order of a triangulated category is one larger than the minimum of the
n-orders of all objects of the form K/n.
r Let S ⊆ T be a full triangulated subcategory and Y an object of S. Induc-
tion on k shows that if n -ordT (Y ) ≥ k, then n -ordS (Y ) ≥ k. Thus we
have n -ordS (Y ) ≥ n -ordT (Y ). In the special case of a zero object we get
n -ord(S) ≥ n -ord(T ).
r Suppose that T is a Z[1/n]-linear triangulated category, i.e., multiplication
by n is an isomorphism for every object of T . Then K/n is trivial for every
object K and thus T has infinite n-order. If on the other hand Y is non-trivial,
then n -ord(Y ) = 0.
r If every object of T has positive n-order, then n · Y = 0 for all objects
Y and so T is a Z/n-linear triangulated category. Suppose conversely
that T is a Z/n-liear triangulated category. Then induction on k shows
that n -ord(Y ) ≥ k for all objects Y , and thus every object has infinite
n-order.
404 Stefan Schwede
The last two items show that the n-order is useless if T is a k-linear triangu-
lated category for some field k, since then every n ∈ Z is either zero or a unit
in k.
The results which we have obtained so far can be rephrased using the notion
of n-order: if T is algebraic, then Propositions 1 and 7 show that for every
object X, the object X/n always has n-order at least 2; we will improve this in
Theorem 16 below. Propositions 4 and 11 show that in the Spanier-Whitehead
category, the Moore spectrum S/2 has 2-order 0 and S/3 has 3-order 1; we
generalize this to mod-p Moore spectra in Theorem 17 below. If T is topological
then for every object X, the object X/3 has 3-order at least 1, by Proposition 5.
If T is topological and p is a prime ≥ 5, then for every object X, the object
X/p has p-order at least 2, by Proposition 12.
The following theorem generalizes Propositions 1 and 7.
Theorem 16. Let T be an algebraic triangulated category and X an object of

T . Then for any n ≥ 2, the object X/n has infinite n-order. In particular, every
algebraic triangulated category T has infinite n-order.
Proof. (Sketch) The assumption that T is algebraic gives another piece of

extra structure: for any integer n there exists a triangulated category T /n and
an adjoint pair of exact functors ρ∗ : T −→ T /n and ρ ∗ : T /n −→ T such
that for every object X of T there exists a distinguished triangle
n· η
X −→ X −→ ρ ∗ (ρ∗ X) −→ X[1]
where η is the unit of the adjunction.

We do not construct T /n in general here, but content ourselves with an
example which gives the main idea. If T = D(A) is the derived category of a
differential graded ring A, then we can take T /n as the derived category of the
differential graded ring A ⊗ Z/n, where Z/n is a flat resolution of the ring Z/n,
for example the exterior algebra over Z on a generator x of dimension 1 with
differential dx = n. The adjoint functor pair (ρ∗ , ρ ∗ ) is derived from restriction
and extension of scalars along the morphism A −→ A ⊗ Z/n.
Now we exploit the extra structure to prove the theorem. Since X/n is
isomorphic to ρ ∗ (ρ∗ X) it is enough to show that for every object Z of T /n and
all k ≥ 0 the object ρ ∗ Z has n-order greater or equal to k.
We proceed by induction on k; for k = 0 there is nothing to prove. Sup-
pose we have already shown that every ρ ∗ Z has n-order greater or equal
to k − 1 for some positive k. Given a morphism f : K −→ ρ ∗ Z in T we
can consider its adjoint fˆ : ρ∗ K −→ Z in T /n; if we apply ρ ∗ we obtain an
extension ρ ∗ (fˆ) : K/n ∼= ρ ∗ (ρ∗ K) −→ ρ ∗ Z of f . We choose a cone of fˆ, i.e.,
a distinguished triangle
fˆ
ρ∗ K −→ Z −→ C(fˆ) −→ ρ∗ K[1]
in T /n. Since ρ ∗ is exact, ρ ∗ C(fˆ) is a cone of the extension ρ ∗ (fˆ) in T . By

induction, ρ ∗ C(fˆ) has n-order greater or equal to k − 1, which proves that ρ ∗ Z
has n-order greater or equal to k.
The next theorem generalizes Propositions 4 and 11 and shows that topo-
logical triangulated categories behave quite differently from algebraic ones.
Theorem 17. Let p be a prime. Then in the Spanier-Whitehead category,

the mod-p Moore spectrum S/p has p-order p − 2. Moreover, the Spanier-
Whitehead category, has p-order p − 1.
The proof of Theorem 17 has two parts. One ingredient is a general state-
ment about p-orders in topological triangulated categories T which generalizes
Propositions 5 and 12: for any object X and prime p, the object X/p has p-
order greater or equal to p − 2. The proof of this result uses the concept and
properties of a coherent action of a mod-p Moore space on an object of a model
category, see Section 2 of [Sch]. It follows that any topological triangulated
category (such as the Spanier-Whitehead category) has p-order at least p − 1.
The second ingredient of Theorem 17 is the proof that the mod-p Moore
spectrum S/p has p-order at most p − 2. This uses mod-p cohomology oper-
ations and serious calculational input; in particular, the proof depends on van-
ishing results, due to other people, about the p-primary components of the
stable stems in specific dimensions. Proposition 11 gives the flavor of the proof
which, however, for general primes p is more involved. I plan to give a detailed
proof elsewhere.
As we just mentioned, every topological triangulated category has p-order
at least p − 1. This leave us with the following question, which generalizes
Problems 6 and 13
Open problem 18. Let p be a an odd prime. Does there exist a triangulated
category whose p-order is strictly less than p − 1?
More generally we can ask which values the n-orders of triangulated cate-
gories can take.
Now that we have discussed torsion phenomena which can distinguish alge-
braic from topological triangulated categories, it is natural to ask whether there
are any differences between topological and algebraic triangulated categories
if all primes are invertible, i.e., in Q-linear triangulated categories. The n-order
406 Stefan Schwede
is rationally a useless invariant since Q-linear triangulated categories have infi-

nite n-order for all n. Similarly, the smash product pairing of a topological
triangulated category with SW gives no extra information for Q-linear trian-
gulated categories since the chain functor C∗ : SW −→ Db (Z) becomes an
equivalence of categories when rationalized (both sides are in fact rationally
equivalent to the category of finite dimensional graded Q-vector spaces).
It turns out that rationally the notions of algebraic and topological triangu-
lated categories essentially coincide. Under some mild technical assumptions
and cardinality restriction, every Q-linear topological triangulated category is
algebraic. More precisely, a theorem of Shipley [Sh, Cor. 2.16] says that every
Q-linear spectral model category (a special kind of stable model category
which is enriched over the stable model category of symmetric spectra) with
a set of compact generators is Quillen-equivalent to dg-modules over a certain
differential graded Q-category. Thus every triangulated category equivalent to
the homotopy category of a stable model category of this kind is algebraic. I
think that the assumption ‘spectral’ is merely of a technical nature and will
eventually be removed. Similarly, I expect that the assumption of a ‘set of
compact generators’ can be relaxed to ‘well generated’ at the price of allowing
localizations of module categories over a differential graded Q-category, along
the lines of the papers [Po] and [He]. At present, I do not know of a Q-linear
triangulated category which is not topological.
Acknowledgement: I would like to thank Andreas Heider for various helpful
comments on this paper.
References
[Ad] J. F. Adams, Stable homotopy and generalised homology. Chicago Lectures
in Mathematics. University of Chicago Press, Chicago, Ill.-London, 1974.
x+373 pp.
[BF] A. K. Bousfield, E. M. Friedlander, Homotopy theory of -spaces, spectra,
and bisimplicial sets. Geometric applications of homotopy theory (Proc.
Conf., Evanston, Ill., 1977), II Lecture Notes in Math., vol. 658, Springer,
Berlin, 1978, pp. 80–130.
[Du] D. Dugger, Spectral enrichments of model categories. Homology, Homotopy
and Applications 8 (2006), 1–30.
[EKMM] A. D. Elmendorf, I. Kriz, M. A. Mandell, J. P. May, Rings, modules, and alge-
bras in stable homotopy theory. With an appendix by M. Cole, Mathematical
Surveys and Monographs, 47, American Mathematical Society, Providence,
RI, 1997, xii+249 pp.
[He] A. Heider, Two results from Morita theory of stable model categories.
!%!%!%!%
[Ho1] M. Hovey, Model categories. Mathematical Surveys and Monographs,

vol. 63, American Mathematical Society, Providence, RI, 1999, xii+209 pp.
[Ho2] M. Hovey, Model category structures on chain complexes of sheaves. Trans.
Amer. Math. Soc. 353 (2001), 2441–2457.
[Ho3] M. Hovey, Spectra and symmetric spectra in general model categories. J.
Pure Appl. Alg. 165 (2001), 63-127.
[HSS] M. Hovey, B. Shipley, J. Smith, Symmetric spectra. J. Amer. Math. Soc. 13
(2000), 149–208.
[Ke] B. Keller, On differential graded categories. Proceedings of the International
Congress of Mathematicians, Madrid, Spain, 2006, vol. II, European Math-
ematical Society, 2006, pp. 151–190.
[Ly] M. Lydakis, Simplicial functors and stable homotopy theory. Preprint (1998).

[MMSS] M. A. Mandell, J. P. May, S. Schwede, B. Shipley, Model categories of
diagram spectra. Proc. London Math. Soc. 82 (2001), 441–512.
[Mar] H. R. Margolis, Spectra and the Steenrod algebra. Modules over the Steenrod
algebra and the stable homotopy category, North-Holland Mathematical
Library 29, North-Holland Publishing Co., Amsterdam-New York, 1983,
xix+489 pp.
[MSS] F. Muro, S. Schwede, N. Strickland, Triangulated categories without models.
Invent. Math. 170 (2007), 231-241.
[Ne] A. Neeman, Triangulated Categories, Annals of Mathematics Studies,
vol. 148, Princeton University Press, Princeton, NJ, 2001.
[Po] M. Porta, The Popescu-Gabriel theorem for triangulated categories.
!%!"##'(
[Q] D. G. Quillen, Homotopical algebra. Lecture Notes in Math. 43, Springer-
Verlag, 1967.
[Sch] S. Schwede, The stable homotopy category is rigid. Annals of Math. 166
(2007), 837-863.
[SS] S. Schwede, B. Shipley, Stable model categories are categories of modules.
Topology 42 (2003), no. 1, 103–153.
[Sh] B. Shipley, H Z-algebra spectra are differential graded algebras. American
J. Math. 129 (2007) 351-379.
[SW] E. Spanier, J. H. C. Whitehead, A first approximation to homotopy theory.
Proc. Nat. Acad. Sci. USA 39 (1953), 655-660.
[To] H. Toda, On spectra realizing exterior parts of the Steenrod algebra. Topol-
ogy 10 (1971), 53–65.
[Vo] R. Vogt, Boardman’s stable homotopy category. Lecture Notes Series, No.
21 Matematisk Institut, Aarhus Universitet, Aarhus 1970 i+246 pp.
Mathematisches Institut, Universität Bonn, Germany

E-mail address:

Derived categories of coherent sheaves on
algebraic varieties
yukinobu toda
Abstract. In this article, we give the introduction of the recent developments
on the derived categories of coherent sheaves on algebraic varieties. We also
introduce the notion of stability conditions on triangulated categories in the sense
of T. Bridgeland.
1. Introduction
The notion of derived category of coherent sheaves was first introduced in [24]
in order to formulate the Grothendieck duality theorem. It is a category whose
objects are complexes of coherent sheaves, and has a structure of a triangulated
category. Recently it has been observed that the derived category represents
several interesting symmetries, which seems impossible without the notion
of derived categories, e.g. McKay correspondence [16], Homological mirror
symmetry [43], etc. Now derived categories are a very popular area with interac-
tions with many other subjects including non-commutative algebra, birational
geometry, symplectic geometry and string theory. In this article, we give an
introduction of the recent results on these topics.
The content of this article is as follows. In Section 2, we give the basic
notions concerning derived categories, and propose some fundamental prob-
lems. In Section 3, we discuss the relationship between derived category
and birational geometry. In Section 4, we discuss the symmetries between
derived categories of coherent sheaves and that of module categories of
some non-commutative algebras. In Section 5, we introduce the notion of
stability conditions on triangulated categories, defined by T. Bridgeland,
and see how this notion explains the several symmetries we discuss in this
article.
Notation and Convention. All the varieties are defined over C. For a variety
X, we denote by Coh(X) the abelian category of coherent sheaves on X. When
X is smooth, we denote by ωX := ∧dim X #X the canonical line bundle, and
KX is the divisor on X such that OX (KX ) = ωX . For a C-vector space V , we
denote by V ∗ its dual vector space.
408
2. Derived categories and Fourier-Mukai transforms

2.1. Derived category of coherent sheaves
For a variety X, its bounded derived category
D(X) := D b Coh(X),
is defined as the localization of the category of bounded complexes of coherent

sheaves with respect to the class of quasi-isomorphisms. So any object in D(X)
is represented by a bounded complex,
· · · → 0 → F m → F m+1 → · · · → F n−1 → F n → 0 → · · · ,
for F i ∈ Coh(X), and a morphism from F • to G • is represented by a diagram,
H• C
{{ CC g
f {
{ CC
{{ CC
}{{ C!
F• G•,
such that f is a quasi-isomorphism and g is a morphism of chain complexes.

The derived category D(X) is no longer an abelian category, but it has a structure
of a triangulated category, i.e.
r There is a shift functor denoted by [1].
r There is a notion of distinguished triangles, F1 → F2 → F3 → F1 [1].
Also these notions should satisfy some axioms. (cf. [24].) On the category
D(X), the shift functor [1] is given by shifting the degree of the complex, and
distinguished triangles are obtained by taking mapping cones of the morphisms
of chain complexes.
Note that the abelian category Coh(X) is regarded as the full subcategory of
D(X), by identifying the objects in Coh(X) with the complexes concentrated
in degree zero. For an object E ∈ D(X), we denote by Hi (E) ∈ Coh(X) the
i-th cohomology of the complex E. The support of E ∈ D(X) is defined by

Supp(E) = Supp Hi (E) ⊂ X.
i
For E, F ∈ D(X) and i ∈ Z, we set
Homi (E, F ) := Hom(E, F [i]).
Here we remark that if E and F are contained in Coh(X), then Homi (E, F )
coincides with the usual Ext-group Exti (E, F ). If X is smooth and projective,
410 Yukinobu Toda
it is well known that the vector space

Homi (E, F )
i∈Z
is finite dimensional. Hence the the following number makes sense,

χ (E, F ) = (−1)i dim Homi (E, F ) ∈ Z.
i∈Z
Now we introduce some functors between derived categories. For the precise
definitions and details, see [24]. Suppose X is smooth and Y is another smooth
variety. Let f : X → Y be a projective morphism. Then we can define the
derived push-forward and derived pull-back,
Rf∗ : D(X) → D(Y ), Lf ∗ : D(Y ) → D(X).
The functor Rf∗ is defined without the assumption that Y is smooth. However
Lf ∗ is not defined if Y is singular. In fact Lf ∗ F for F ∈ D(Y ) may be
unbounded below, hence one has to replace D(X) by the unbounded (or bounded
above) derived category of coherent sheaves. Also for an object E ∈ D(X), we
can define the derived homomorphism and derived tensor product,
L
RHom(E, ∗) : D(X) → D(X), ∗ ⊗ E : D(X) → D(X).
Suppose Y is an affine variety Spec R, not necessary smooth. Then D(Y ) is

naturally identified with the bounded derived category of finitely generated
R-modules, denoted by D(R). For E ∈ D(X), we define the functor,
R Hom(E, ∗) := Rf∗ ◦ RHom(E, ∗) : D(X) −→ D(R).
Note that we have
Hi (R Hom(E, F )) = Homi (E, F ),
as R-modules for F ∈ D(X). In this article we sometimes omit L if the functors

we derive are exact. For example, if f is flat we write Lf ∗ = f ∗ .
2.2. Fourier-Mukai transforms

Let X be a smooth projective variety. It is known that the abelian category
Coh(X) determines X itself. (For example see [39, Theorem 3.2].) On the other
hand, the derived category D(X) does not necessary determine X itself. The
reason behind this is that we have additional correspondences called Fourier-
Mukai transforms.
Definition 2.1. Let X and Y be smooth projective varieties. For an object

P ∈ D(X × Y ), define the functor PX→Y : D(X) → D(Y ) by the formula,
L
P ∗
X→Y (E) = RpY ∗ (pX E ⊗ P),
for E ∈ D(X). Here pX , pY are projections from X × Y onto corresponding

factors. If PX→Y gives an equivalence, it is called a Fourier-Mukai transform.
The object P is called a kernel of the functor PX→Y .
The functors of Fourier-Mukai type are not special equivalences. In fact

Orlov [52] proved the following.
Theorem 2.2. (Orlov [52]) Let : D(X) → D(Y ) be an equivalence of C-

linear triangulated categories. Then there exists an object P ∈ D(X × Y ) such
that is isomorphic to the functor PX→Y . Moreover P is uniquely determined
up to isomorphism.
Here we give some examples of Fourier-Mukai transforms.
Example 2.3. (Mukai [48], [49]) (i) Let A be an abelian variety, i.e. A is
a complex torus Cn / embedded into a projective space. Here ⊂ R2n is a
lattice of rank 2n. Let Â = Pic0 (A) be the moduli space of line bundles on
A which are topologically isomorphic to OA . It is known that Â is also an
abelian variety of the same dimension, and called dual abelian variety. Let
U ∈ Pic(A × Â) be the universal family. Then the functor
U
Â→A
: D(Â) −→ D(A),
is a Fourier-Mukai transform. See [48] for the detail. When dim A ≥ 2, the
dual Â is not necessary isomorphic to A. This means that the derived category
D(A) does not determine the original variety A.
(ii) An analogue of (i) for K3 surfaces was studied in [48]. Let X be a K3
surface, i.e. simply connected algebraic surface with ωX trivial. Let M be a
connected component of the moduli space of stable sheaves on X, which is two
dimensional, projective, and has a universal family E ∈ Coh(X × M). Then the
functor
E
M→X : D(M) −→ D(X),
is a Fourier-Mukai transform.
412 Yukinobu Toda
2.3. Fourier-Mukai partners

As we observed in Example 2.3, for a given smooth projective variety X there
may be several varieties Y which are derived equivalent to X. So we have the
following natural problem.
Problem 2.4. When there exists a Fourier-Mukai transform,
: D(Y ) −→ D(X),
is there geometric relationship between X and Y ?
If Y is derived equivalent to X, then it is called a Fourier-Mukai partner

of X. The Problem 2.4 was first considered by Bondal and Orlov [8], under a
geometrically strong assumption that KX or −KX is ample.
Theorem 2.5. (Bondal-Orlov [8]) Assume KX or −KX is ample. Then any

Fourier-Mukai partner of X is isomorphic to X itself.
The class of varieties X with KX or −KX ample includes the hyper-

surfaces of degree d in Pn with d = n + 1. One of the key points in
studying Fourier-Mukai partners is to investigate the several categori-
cal invariants of derived categories, such as the Serre functor we define
below.
Definition 2.6. Let T be a C-linear triangulated category. An equivalence

S : T → T as triangulated categories is called a Serre functor if there exists a
bifunctorial isomorphism
Hom(E, F ) → Hom(F, S(E))∗
for E, F ∈ T .
It is shown in [8, Proposition 1.5] that if a Serre functor exists, then it is

unique up to canonical isomorphism. When T = D(X) for a smooth projective
variety X, then Serre duality implies that
SX (E) = E ⊗ ωX [dim X],
for E ∈ D(X) gives a Serre functor on D(X).

Another key point is to find the objects of the form Ox for closed points
x ∈ X from the categorical data of D(X).
Definition 2.7. [8, Definition 2.1] An object P ∈ D(X) is called a point object
of codimension s if it satisfies,
(i) SX (P ) ∼
= P [s],
(ii) Hom<0 (P , P ) = 0,
(iii) Hom0 (P , P ) = C.
Now the idea of the proof of Theorem 2.5 is explained as follows. Note
that the notion of point objects are determined by only the data of D(X) as a
triangulated category. In [8, Proposition 2.2], it is shown that if KX or −KX
is ample, then any point objects are of the form Ox [r] for some closed point
x ∈ X and r ∈ Z. Thus one can reconstruct the notion of “point” from the
data of D(X) as a triangulated category, which enables us to reconstruct X
from D(X).
On the other hand if KX is trivial, then the condition (i) of Definition 2.7 does
not say anything so there may be other point objects, and non-trivial Fourier-
Mukai partners. In fact the examples given in Example 2.3 are such cases.
At this time, very few is known about Fourier-Mukai partners. Other than
Theorem 2.5, we know the following cases:
(i) The case of dim X ≤ 1.

In this case any Fourier-Mukai partner of X is known to be isomorphic to
X. In fact if X is not an elliptic curve, this follows from Theorem 2.5. If X is
an elliptic curve, see [25, Theorem 5.1].
(ii) The case of dim X = 2.

When X is minimal i.e. KX · C ≥ 0 for any curve C ⊂ X, it is shown in [17]
that any Fourier-Mukai partner of X is isomorphic to X itself, except X is a K3
surface or an abelian surface. When X is a K3 surface or an abelian surface,
any Fourier-Mukai partner is isomorphic to the moduli space of stable sheaves
on X, as in Example 2.3 (ii). When X is not minimal, then any Fourier-Mukai
partner of X is isomorphic to X itself except X is a relatively minimal rational
elliptic surface [37].
(iii) The case of dim X = 3.

When κ(X) = 3, any Fourier-Mukai partner is obtained as a sequence of
flops [37]. (See (1) in the next paragraph for the definition of κ(X), and Def-
inition 3.4 for flops.) When κ(X) = 2, any Fourier-Mukai partner is obtained
as a relative Jacobian up to flops [60].
414 Yukinobu Toda
(iv) The case that X is an abelian variety A.

In this case, any Fourier-Mukai partner B is also an abelian variety, and an
abelian variety B is a Fourier-Mukai partner of A if and only if there exists a
symplectic isomorphism [53],
Â × A −→ B̂ × B.
As is observed in the above cases, the known examples of Fourier-Mukai

partners are geometrically quite similar properties.
2.4. Several categorical invariants

It seems that the classification problem of Fourier-Mukai partners are quite
difficult, especially when Kodaira dimension of the variety in question is near
to zero. On the other hand we know several categorical invariants of D(X) as
a triangulated category, and such invariants would help us how Fourier-Mukai
partners look like. Let X and Y be smooth projective varieties and assume there
exists an equivalence of derived categories : D(X) → D(Y ).
(i) Invariance of dimensions. We have dim X = dim Y . (cf. [37, Theorem

1.4].)
(ii) Invariance of cohomology groups. In [52], it is shown that induces

an isomorphism of vector spaces, φ : H ∗ (X, Q) → H ∗ (Y, Q) such that the
following diagram commutes,
D(X) / D(Y )
√ √
ch(∗) tdX ch(∗) tdY

H ∗ (X, Q)
φ
/ H ∗ (Y, Q).
The commutativity of the above diagram is due to Grothendieck Riemann Roch

theorem. The isomorphism φ is given by the correspondence of the following
algebraic cycle,
8 8
∗
pX tdX · ch(P) · pY∗ tdY ∈ H ∗ (X × Y, Q),
where P ∈ D(X × Y ) is a kernel of . Note that φ does not necessary preserve

the degree of the cohomology ring. In [52], the existence of φ is a crucial point
in reducing the classification problem of Fourier-Mukai partners of K3 surfaces
to Torelli type problem. Also Orlov [54] conjectures the relationship between
derived equivalence and Chow motives.
(iii) Invariance of canonical rings. In [18], [60], it is shown that induces

an isomorphism of graded rings

H 0 (X, mKX ) −→ H 0 (Y, mKY ).
m∈Z m∈Z
In particular we have κ(X) = κ(Y ). (Also see [37].) Here κ(X) is the Kodaira
dimension, defined by the number k ∈ Z such that
H 0 (X, mKX ) ∼ amk , (m 0) (1)
for some a > 0 if H 0 (X, mKX ) is non-zero for some m > 0. If H 0 (X, mKX ) =
0 for any m > 0, κ(X) is defined to be −∞. The idea of [60] is to reduce the
classification problem of Fourier-Mukai partners to the case of Kodaira dimen-
sion zero, by comparing the derived categories of the fibers of the following
rational maps called Iitaka fibrations,

X Proj H 0 (X, mKX ), Y Proj H 0 (X, mKY ).
m≥0 m≥0
(iv) Invariance of Hochschild cohomologies. Let H H N (X) and H T N (X)

be
X×X (OX , OX ),

H H N (X) := ExtN (2)
9
j
H T N (X) := H i (X, TX ). (3)
i+j =N
Here X ⊂ X × X is a diagonal, and H H N (X) is called Hochschild cohomol-

ogy. Then the equivalence induces an isomorphism [18],
∼
=
φH : H H N (X) −→ H H N (Y ).
On the other hand, we have the following isomorphism called HKR isomor-
phism [26], [45], [58], [67].
∼
=
IH KR : H H N (X) −→ H T N (X).
Combined with φH and IH KR , we obtain the isomorphism,
∼
=
φT : H T N (X) −→ H T N (Y ).
(v) Invariance of deformation theories. As a special case of (iv), the

following vector space is preserved by the derived equivalence,
9
2
H T 2 (X) = H 2 (X, OX ) ⊕ H 1 (X, TX ) ⊕ H 0 (X, TX ).
416 Yukinobu Toda
The above space parameterizes the first order deformations of the abelian cate-
gory Coh(X). (See [45], [46], [64].) In fact for u = (α, β, γ ), β corresponds to
the usual deformation of X as a scheme, γ corresponds to a non-commutative
deformation, and α is a gerby deformation. One can construct the C[ε]/(ε 2 )-
linear abelian category Coh(X, u), which is a mixture of the above deforma-
tions. Then the equivalence extends to an equivalence [64],
†
: D b Coh(X, u) −→ D b Coh(Y, v),
where v = φT (u) ∈ H T 2 (Y ). Extending † to infinite order deformations is

problematic. At this time, † is known to be extended to infinite order defor-
mations when the equivalence is given as in Example 2.3 (i). (See [4].) We
expect that the deformation methods will enable us to find more interesting
Fourier-Mukai dualities.
2.5. The group of autoequivalences

The next problem is similar to Problem 2.4.
Problem 2.8. Compute the group of autoequivalences Auteq D(X).
When KX or −KX is ample, Bondal and Orlov [8] showed (by the same idea
of Theorem 2.5) that the group Auteq D(X) is generated by trivial autoequiva-
lences, that is the group of automorphisms Aut(X), tensoring line bundles, and
the shift functor [1]. Again it seems that Auteq D(X) is more difficult when
the Kodaira dimension of X is near to zero. Non-trivial autoequivalences are
provided by the spherical objects below.
Definition 2.9. An object E ∈ D(X) is said to be a spherical object if and

only if

r Homi (E, E) = C (i = 0, dim X),
0 (otherwise),
r E ⊗ ωX ∼ = E.
Under mirror symmetry, a spherical object should correspond to a

Lagrangian sphere, which has an associated symplectic isomorphism called
Dehn twist [56]. Thus one expects that a spherical object associates a corre-
sponding autoequivalence on D(X). The answer was provided by Seidel and
Thomas [57].
Theorem 2.10. (Seidel-Thomas [57]) For a spherical object E ∈ D(X), there

exists an autoequivalence TE ∈ Auteq D(X) such that for any F ∈ D(X), one
has a distinguished triangle,
R Hom(E, F ) ⊗ E −→ F −→ TE (F ).
For a spherical object E, its associated autoequivalence TE is called spherical

twist. There are certain generalizations of this twist [27], [61].
The recent work of Bridgeland [15] relates the computation of Auteq D(X)
to the topology of the space of stability conditions on D(X). We discuss this
in Section 5. Here we introduce two important computations of the group
Auteq D(X).
Autoequivalences on abelian varieties. Let A be an abelian variety and Â

be its dual abelian variety. We can write a morphism f : A × Â → A × Â as a
matrix,

α β
f = ,
γ δ
and define f: A × Â → A × Â to be

δ̂ −γ̂
f = .
−β̂ α̂
Then we introduce the group U (A × Â) as follows,
U (A × Â) := {f ∈ Aut(A × Â) | f = f −1 }.
In the paper [53], Orlov showed the following.
Theorem 2.11. (Orlov [53]) For an abelian variety A, there exists the following
exact sequence,
0 −→ Z ⊕ (A × Â) −→ Auteq D(X) −→ U (A × Â) −→ 1.
Autoequivalences on An -configurations of (−2)-curves. Another impor-

tant case, the group of autoequivalences on An -configurations of (−2)-curves,
was computed by Ishii and Uehara [29]. Let Y be an An -singularity,
Y = Spec C[x, y, z]/(xy + zn+1 ),
and f : X → Y be the minimal resolution. Note that the exceptional locus is a

chain of rational curves,
Z = C1 ∪ · · · ∪ Cn , Ci ∼
= P1 .
418 Yukinobu Toda
Under this situation, Ishii and Uehara [29] considered the triangulated category,
DZ (X) = {E ∈ D(X) | Supp(E) ⊂ Z},
and computed the group,
AuteqFM (DZ (X)) = { ∈ Auteq(DZ (X)) |

∼
= PX→X with Supp(P) ⊂ X ×Y X}.
Note that OCi (−1), ωZ are spherical objects, so there are associated twists
TOCi (−1) , TωZ . We set
B := TOCi (−1) , TωZ | 1 ≤ i ≤ n ⊂ AuteqFM (DZ (X)).
Theorem 2.12. (Ishii-Uehara [29]) We have
AuteqFM (DZ (X)) = (B, Pic(X) Aut(X)) × Z.
Here Z is generated by the shift functor [1].
Furthermore Ishii, Ueda and Uehara [28] determined the group structure of
B using the space of stability conditions.
3. Derived category and birational geometry

In this section, we discuss the relationship between derived category and bira-
tional geometry. Recall that minimal model program, referred as MMP, is a
program to find a good birational model (minimal model or Fano fiber space)
by contracting extra rational curves [44]. For a given variety X, MMP suggests
that there is a sequence,
X = X0 X1 X2 · · · Xm ,
such that each Xi Xi+1 is a birational morphism which contracts a divisor,
or a certain birational map called flip. The output Xm is either a minimal model
(i.e. KXm · C ≥ 0 for any curve C ⊂ Xm ) or has a fiber space structure Xm → Z
such that the general fiber is a Fano manifold. The purpose of this section is to
observe that MMP is more natural in the context of the “space” stated in the
introduction.
3.1. Semiorthogonal decomposition

As a preparation, we introduce the notion of semiorthogonal decomposition.
Definition 3.1. Let T be a triangulated category and i : C ⊂ T be a full

triangulated subcategory. We say C is right admissible if the inclusion i admits
a right adjoint p : T → C.
One can show the following: for a full subcategory C ⊂ T , it is right admis-
sible if and only if for E ∈ T , there is a distinguished triangle,
E1 −→ E −→ E2 ,
with E1 ∈ C and E2 ∈ C ⊥ = {F ∈ T | Hom(C, F ) = 0}.
Definition 3.2. Let C1 , · · · , Cm be full triangulated subcategories of T , and

Ti ⊂ T the smallest triangulated subcategory which contains C1 , · · · , Ci . Sup-
pose that
r T = Tm .
r Ci is a right admissible subcategory of Ti .
In this case, we write
T =< C1 , C2 , · · · , Cm >,
and it is called a semiorthogonal decomposition of T .
3.2. Blow-up formula of derived categories

The behavior of derived categories under birational transformations was first
studied by Bondal and Orlov [7]. Let X be a smooth projective variety and
Z ⊂ X a smooth closed subscheme of codimension r. Let f : Y → X be
the blow-up along Z, and E ⊂ Y the exceptional divisor of f . We have the
following diagram,
E
i /Y
p f

Z / X.
Let OE (1) ∈ Pic(E) be the tautological line bundle of the projective bundle
E = P(NZ/X ). Bondal and Orlov [7] showed the following.
Theorem 3.3. (Bondal-Orlov [7]) The functors,
Lf ∗ : D(X) → D(Y ), i∗ ◦ (p∗ (∗) ⊗ OE (j )) : D(Z) → D(Y ),

420 Yukinobu Toda
are fully faithful for −r + 1 ≤ j ≤ −1. We have the following semiorthogonal

decomposition,
D(Y ) =< D(E)−r+1 , · · · , D(E)−1 , Lf ∗ D(X) >,
where D(E)j is the image of i∗ ◦ (p ∗ (∗) ⊗ OE (j )).
Theorem 3.3 asserts that the derived category will be bigger by taking blow-
ups, and the difference is described by the derived category of the exceptional
locus.
3.3. Derived equivalence under birational transformations

Next we discuss the derived equivalence between birational minimal models.
First we recall the notion of flip and flop. (See [42].)
Definition 3.4. Let X and X† be smooth quasi projective varieties and assume
there exist projective birational morphisms,
f g
X −→ Y ←− X† ,
to a (singular) variety Y , which satisfy the following.

r f and g are isomorphisms in codimension one.
r The relative Picard numbers of f and g are one respectively.
r For a f -ample divisor D on X, let D † = φ∗ D be the strict transform on X†
with respect to the birational map φ = g −1 ◦ f : X X† . Then −D † is
g-ample.
If furthermore −KX is ample over Y (resp numerically zero over Y ), then φ is

called flip (resp flop).
Remark 3.5. Usually flips and flops are defined over varieties with terminal
singularities or pairs (X, B) with klt singularities, because MMP works in the
category of varieties or pairs with such singularities. (See [44].) However we
omit this general treatment in order to simplify the argument.
Example 3.6. Let Y be a conifold singularity,
Y = Spec C[x, y, z, w]/(xy − zw).
Let f : X → Y and g : X † → Y be the blowing ups at the ideals (x, z) ⊂ OY ,

f g
(x, w) ⊂ OY respectively. Then X → Y ← X† satisfies Definition 3.4 with KX
numerically zero over Y . It is called Atiyah flop.
In [7], Bondal and Orlov studied the behavior of the derived categories
under certain special flips and flops. Let f : X → Y be as in Definition 3.4
with dim X = n. Suppose the exceptional locus E is isomorphic to Pr with
normal bundle OE (−1)⊕n−r . Then we have the diagram,
ZA
~~ AA
p
~~ AAq
~~ AA
~
~ A
X? X†
?? }}
??
? }}}g
f ??
~}
}
Y,
such that p is a blow up along E, q is also a blow up along the subscheme E † ⊂

X† which is isomorphic to Ps with r + s = n − 1. If r > s (resp r = s), the
above diagram gives a flip (resp flop). Using the blow-up formula Theorem 3.3,
Bondal and Orlov [7] showed the following.
Theorem 3.7. (Bondal-Orlov [7]) The functor
Rp∗ ◦ Lq ∗ : D(X† ) −→ D(X),
is fully faithful if r ≥ s and equivalence if r = s.
It is known that any birational minimal models are connected by flops [35].
Hence Theorem 3.7 leads us to the following conjecture.
Conjecture 3.8. Let X and Y be birationally equivalent minimal models. Then

there is an equivalence of derived categories, D(X) → D(Y ).
Theorem 3.3, Theorem 3.7 and Conjecture 3.8 indicate the following. One of
the features of MMP in dimension bigger or equal to three is that the minimal
model is not necessary unique. However if Conjecture 3.8 is true, then it is
unique on the level of derived categories. Hence if we have some kind of nice
geometry of “spaces” as in the introduction, and an appropriate theory of the
analogue of minimal model theory, then the corresponding minimal model must
be unique. Furthermore, because derived categories will be smaller by blow
downs and flips by Theorem 3.3, Theorem 3.7, the minimal model theory in
the “space” should correspond to “minimizing” derived categories. This is the
reason why the geometry needed for string theory is also natural in birational
geometry, and hopefully it will enable string theory to interact with birational
geometry. However there are several technical problems in realizing this story,
and it is only just a philosophy at this time.
422 Yukinobu Toda
3.4. Perverse t-structures and flops

In dimension three, Conjecture 3.8 was proved by Bridgeland [11] using the
technique of t-structures.
Definition 3.9. Let T be a triangulated category. A full subcategory T ≤0 ⊂ T

is called a t-structure if
r We have T ≤0 [1] ⊂ T ≤0 .
r For any E ∈ T , there is a distinguished triangle,
τ≤0 E −→ E −→ τ≥1 E,
such that τ≤0 E ∈ T ≤0 and τ≥1 E ∈ T ≥1 . Here T ≥1 := {F ∈ T |

Hom(T ≤0 , F ) = 0}.
The subcategory
T ≤0 ∩ T ≥1 [1] ⊂ T ,
is called the heart of the t-structure. It is an abelian subcategory [3].
Example 3.10. For a variety X, let D ≤0 (X) be
D ≤0 (X) = {E ∈ D(X) | Hi (E) = 0 for i > 0}.
Then it is a t-structure on D(X). Its heart is given by Coh(X) ⊂ D(X), con-

centrated in degree zero. The t-structure obtained in this way is called standard
t-structure.
Let f : X → Y be as in Definition 3.4 with dim X = 3. Bridgeland consid-

ered the following abelian categories.
Proposition 3.11. ([11]) There are t-structures on D(X) with hearts i Per(X/Y )
for i = −1, 0 such that an object E ∈ D(X) is contained in i Per(X/Y ) if and
only if
r Hj (E) = 0 unless j = −1, 0.
r R 1 f∗ H0 (E) = 0 and f∗ H−1 (E) = 0.
r For any c ∈ Coh(X) with Rf∗ c = 0, we have
Hom(H0 (E), c) = 0 (i = −1),

−1
Hom(c, H (E)) = 0 (i = 0).
An object in −1 Per(X/Y ) ⊂ D(X) is called a perverse sheaf because its

construction is an analogue of the perverse sheaves on the derived category of
constructible sheaves [3]. Bridgeland’s result [11] is the following.
Theorem 3.12. (Bridgeland [11]) For a three dimensional flop, X → Y ←

X† , there exists an equivalence of derived categories,
: D(X† ) −→ D(X),
which takes −1 Per(X† /Y ) to 0 Per(X/Y ). In particular Conjecture 3.8 is true in
dimension three.
Proof. We explain the idea of the proof of Theorem 3.12. By the construction,
the structure sheaf OX is a perverse sheaf. We call an object E ∈ −1 Per(X/Y )
perverse point sheaf if there is a surjection OX E in −1 Per(X/Y ), and E is
numerically equivalent to a skyscraper sheaf Ox for a closed point x ∈ X. Here
we note that any skyscraper sheaf Ox is a perverse sheaf, but the natural mor-
phism OX → Ox is not necessary surjective in −1 Per(X/Y ). Thus the notion
of perverse point sheaves is different from the notion of skyscraper sheaves.
Then Bridgeland constructed the moduli space of perverse point sheaves
P-Hilb(X/Y ) using the method of GIT-quotient. It admits a morphism to Y ,
P-Hilb(X/Y ) E
−→ Supp(Rf∗ E) ∈ Y.
The space P-Hilb(X/Y ) is called perverse Hilbert scheme, and there is a
universal perverse point sheaves,
E ∈ D(X × P-Hilb(X/Y )).
One can see that there exists a unique irreducible component
X† ⊂ P-Hilb(X/Y ) which dominates Y . Then Bridgeland showed that
X† is smooth, X† → Y gives a flop, and the restriction E ◦ = E|X×X† gives the
desired derived equivalence,
E◦
X† →X : D(X† ) −→ D(X),
using the intersection theorem in the theory of commutative algebra. It requires
the condition dim X† ×Y X† ≤ dim X + 1, which holds automatically in the
case of dim X = 3.
It is important to notice that Bridgeland’s method guarantees the existence
of a flop and its smoothness at the same time. The existence of a flop is a
non-trivial problem in birational geometry, so this argument provides a new
viewpoint in birational geometry.
The result of Theorem 3.12 was generalized to flops of certain singular
threefolds with terminal singularities [19], [37]. However it seems difficult to
extend his result to higher dimensional flops. In fact there are examples of
four dimensional flops X X† which does not preserve the smoothness of
varieties. (See [38].) Thus his argument must fail in such cases, because if it
works, then the flop must also be smooth.
424 Yukinobu Toda
3.5. D-equivalence and K -equivalence

Let X and Y be smooth projective varieties. We say X and Y are K-equivalent
if there exists another smooth projective variety Z and a diagram
Z@
@@ q
p @@
@@

X Y,
such that p, q are birational and p ∗ KX = q ∗ KY . It is well known that if X and Y
are birational minimal models, then they are K-equivalent. Also Kawamata [37]
called the two varieties X, Y D-equivalent if Y is a Fourier-Mukai partner of X.
As a generalization of Conjecture 3.8, Kawamata [37] proposed the following
conjecture.
Conjecture 3.13. (Kawamata [37]) For birationally equivalent smooth pro-
jective varieties X and Y , they are K-equivalent if and only if they are D-
equivalent.
Uehara [66] found an example of birationally equivalent varieties X and
Y whose derived categories are equivalent, but they are not K-equivalent.
Thus Conjecture 3.14 is not true in the sense that D-equivalence does not
imply K-equivalence. However there are some results, including Theorem 3.7,
Theorem 3.12, which support the conjecture that K-equivalence implies D-
equivalence. Here we put another evidence for this.
Theorem 3.14. (Kawamata [37], Namikawa [50]) Suppose that X and X †
are related by a Mukai flop X → Y ← X† , i.e. there is a diagram
ZA
AA p
p1
AA 2
AA

X? X†
?? ~
?? ~~
?
f1 ?? ~~~f2
~~
~
Y ,
such that pi is the blowing up of a subvariety Pd ∼

= Ei ⊂ X with normal bundle
equal to #Ei and fi contracts Ei to a point. Then the functor
OX× †
YX
X† →X
: D(X† ) −→ D(X),
gives an equivalence. However the functor Rp2∗ ◦ Lp1∗ does not give an
equivalence.
4. Derived categories and non-commutative algebras

In this section, we discuss symmetries between usual varieties and several
non-commutative algebras.
4.1. Tilting generators

Let X be a smooth variety and R be a noetherian C-algebra. We assume there
exists a projective morphism f : X → Spec R.
Definition 4.1. An object E ∈ D(X) is a tilting generator if the following

conditions are satisfied.
r R Hom(E, E) = Hom0 (E, E).
r If an object F ∈ D(X) satisfies R Hom(E, F ) = 0, then F ∼
= 0.
Suppose E ∈ D(X) is a tilting generator and set A = End(E). Note that A

is a finitely generated R-algebra. Furthermore for F ∈ D(X), the R-module
Hom(E, F ) is regarded as a right A-module in the obvious way. Hence we have
the factorization,
R Hom(E, ∗) : D(X) −→ D(mod A) −→ D(R).

1 2
Here mod A is the abelian category of right A-modules, and the functor 2 is
forgetting the A-module structure via the natural morphism R → A. By abuse
of notation, we also use the notation R Hom(E, ∗) for the functor 1 . As an
analogue of Morita correspondence, we have the following theorem.
Theorem 4.2. Let E ∈ D(X) be a tilting generator, and A = End(E). Then

the following functor
R Hom(E, ∗) : D(X) −→ D(mod A),

L
gives an equivalence. The quasi-inverse is given by M
→ E ⊗ M.
One can consult [25, Theorem 7.6] for the proof of Theorem 4.2.
4.2. Exceptional collections

When R = C, one can construct the tilting generator from an exceptional
collection.
426 Yukinobu Toda
Definition 4.3. For a C-linear triangulated category T , a sequence of objects

E1 , · · · , En is called an exceptional collection if

C (k = 0, i = j ),
Homk (Ei , Ej ) =
0 (k = 0, i = j or i > j ).
It is called strong if furthermore
Homk (Ei , Ej ) = 0, (k = 0, i < j ).
An exceptional collection E1 , · · · , En is called full if the smallest triangulated

subcategory in T which contains E1 , · · · , En is T itself. In this case we write
T =< E1 , · · · , En > .
Suppose a sequence E1 , · · · , En is a full strong exceptional collection on

D(X), where X is a smooth projective variety over C. Then the object

n
E= Ei ,
i=1
is a tilting generator by the definition. Hence there is an equivalence of derived

categories,
R Hom(E, ∗) : D(X) −→ D(mod A),
where A = End(E). This sort of equivalence was first studied by Beilinson [2].
Theorem 4.4. (Beilinson [2]) We have the following description of the derived
category of Pn .
D(Pn ) =< OPn (−n), · · · , OPn (−1), OPn > .
In particular we have the equivalence,
D(Pn ) −→ D(mod A),
where A = End ⊕ni=0 OPn (−i).
Recently Y. Kawamata [40] generalized Theorem 4.4 to smooth toric vari-

eties, using minimal model program for toric varieties.
Theorem 4.5. (Kawamata [40]) Let X be a smooth projective toric variety.

Then D(X) admits a full strong exceptional collection consisting of sheaves.
4.3. McKay correspondence

Let Y be a variety with at most Gorenstein singularities. We say a resolution of
singularities f : X → Y crepant if ωX = f ∗ ωY . Suppose that a group G acts
on a variety X. We denote by DG (X) the derived category of G-equivariant
coherent sheaves on X. The following is the most sophisticated version of
McKay correspondence stated by M. Reid [55].
Conjecture 4.6. Let G ⊂ SL(n, C) be a finite subgroup. Suppose that there

exists a crepant resolution X → Cn /G. Then there exists an equivalence of
derived categories,
D(X) ∼
= DG (Cn ).
When n ≤ 3, there is a positive answer for Conjecture 4.6.
Theorem 4.7. (Bridgeland, King, Reid [16]) Conjecture 4.6 is true when
n ≤ 3.
Proof. We explain the idea of [16], which is similar to the proof of Theo-
rem 3.12. First we consider the G-clusters on C3 . These are G-invariant closed
subschemes Z ⊂ C3 which satisfy OZ ∼ = C[G]. Then consider the moduli
space of G-clusters, denoted by G- Hilb(C3 ). By associating a G-cluster to its
support, one has the map
G- Hilb(C3 ) −→ C3 /G.
Then it has a unique irreducible component X ⊂ G- Hilb(C3 ) which dominates

C3 /G. We have a universal G-clusters Z ⊂ X × C3 , and the diagram,
Z
x GGG
q xxx GGp
xx GG
x GG
|x
x #
XD
DD C3
DD yy
DD yy
f DD yyyπ
" |y
C3 /G.
Then f : X → C3 /G gives the crepant resolution and the transformation
Rq∗ ◦ Lp∗ : D(X) −→ DG (C3 ),
gives an equivalence.
428 Yukinobu Toda
As in the proof of Theorem 3.12, we have to use the intersection theorem to

show that X is smooth and the functor Rq∗ ◦ Lp∗ gives an equivalence. This
is why the above proof works only for dimension not greater than three.
On the other hand, Conjecture 4.6 has been shown to hold in some other
situations. One of them is due to Bezrukavnikov and Kaledin [6].
Theorem 4.8. (Bezrukavnikov, Kaledin [6]) Conjecture 4.6 holds when G

preserves a complex symplectic form on Cn .
Their method is very different from the proof of Theorem 4.7, using charac-
teristic p methods and deformation quantization. Another special case is due
to Kawamata [40].
Theorem 4.9. (Kawamata [40]) Conjecture 4.6 holds when G is an abelian.
4.4. Non-commutative crepant resolutions

Note that the category of G-equivariant sheaves on C3 is equivalent to
the category of the module categories over the smash product algebra
A = C[x, y, z]&G. Thus McKay correspondence is interpreted as a symme-
try between (usual) commutative schemes and non-commutative algebras.
In McKay correspondence, the story starts by giving a non-commutative
algebra A at first, which seems to be a non-commutative analogue of crepant
resolutions. Thus conversely, given a crepant resolution f : X → Y = Spec R
at first, one can ask whether there exists a non-commutative algebra A which
is an analogue of a crepant resolution, and it admits a derived equivalence
D(X) ∼ = D(mod A). M. Van den Bergh [20] introduced the notion of non-
commutative crepant resolution in order to formulate this problem.
Definition 4.10. (M. Van den Bergh [20]) A non-commutative crepant resolu-
tion of R is an homologically homogeneous R-algebra of the form A = End(M)
where M is a reflective R-module.
Here the notion of homologically homogeneous is a non-commutative ana-

logue of smoothness. In [20], M. Van den Bergh explains why the above defi-
nition is interpreted as the notion of crepant resolution in non-commutative
algebras. He also constructed a non-commutative crepant resolution when
dimensions of fibers of f : X → Spec R are less or equal to one.
Theorem 4.11. (M. Van den Bergh [21]) Assume dim f −1 (p) ≤ 1 for any
p ∈ Spec R. Then there exists a tilting generator E ∈ D(X) such that A =
End(E) is a non-commutative crepant resolution. In particular there is a derived
equivalence,
D(X) −→ D(mod A).
Proof. We give one of the constructions of the tilting generator E in [21].

First take a f -ample line bundle L ∈ Pic(X), generated by its global sections.
Because R 1 f∗ L−1 is a finitely generated R-module, there is a surjection R ⊕n
R 1 f∗ L−1 . Then there is a universal extension,
⊕n
0 −→ L−1 −→ E −→ OX −→ 0.
Note that E is a vector bundle on X and satisfies R Hom(E, E) = End(E). Then

E = OX ⊕ E gives a tilting generator.
Example 4.12. Let f : X → Y be as in Example 3.6, and C ∼ = P1 ⊂ X be

the exceptional locus. Then E0 = OX ⊕ OX (−1) and E−1 = OX ⊕ OX (1) are
tilting generators. We have the associated derived equivalences,
i: D(X) −→ D(mod Ai ),
for i = −1, 0, where Ai = End(Ei ). The algebras A−1 , A0 are path algebras
with certain relations corresponding to the quiver given in Figure 4.13. Espe-
cially the category mod Ai has two simple objects, corresponding to the two
vertex in the quiver. In [21], it is shown that i takes i Per(X/Y ) to mod Ai .
The corresponding simple objects in −1 Per(X/Y ) are given by
{OC (−1), OC (−2)[1]} ⊂ 0 Per(X/Y ), {OC (−1)[1], OC } ⊂ −1 Per(X/Y ),
respectively.
Figure 4.13.
(
•h (•
h
In dimension three, M. Van den Bergh [20] also proved the following.
Theorem 4.14. (M. Van den Bergh [20]) Assume that R is three dimen-
sional and has terminal singularities. Then R has a non-commutative crepant
resolution if and only if it has a commutative one. Furthermore any two
crepant resolutions (commutative as well as non-commutative ones) are derived
equivalent.
430 Yukinobu Toda
5. Stability conditions on triangulated categories

In this section, we introduce the notion of stability conditions on triangulated
categories. It was introduced by T. Bridgeland [14] to give the mathematical
framework for M. Douglas’s work on '-stability [22], [23]. Its motivation
comes from physics, however it is also interesting in mathematics, and espe-
cially gives nice pictures for the symmetries we have discussed so far.
5.1. Stability conditions on coherent sheaves on curves

First we recall the notion of stability condition on the abelian category of
coherent sheaves on a smooth projective curve C. For E ∈ Coh(C), we denote
μ(E) ∈ Q ∪ {∞} as follows:

∞ if E is torsion ,
μ(E) =
deg E/ rank E otherwise .
Definition 5.1. An object E ∈ Coh(C) is called semistable if for any subsheaf
F ⊂ E, one has μ(F ) ≤ μ(E).
We have the following two properties for semistable sheaves.
r For any E ∈ Coh(C), there is a unique filtration
0 = E0 ⊂ E1 ⊂ · · · ⊂ En = E, (4)
such that Ai = Ei /Ei−1 is semistable and μ(Ai ) > μ(Ai+1 ) for all i. The
filtration (4) is called Harder-Narasimhan filtration.
r There is a moduli space of semistable sheaves with a fixed rank and a degree.
Let us take E ∈ D(C). Then by taking Harder-Narasimhan filtrations (4) for
Hi (E) ∈ Coh(C), we obtain the sequence,
0 = E0 / E1 / E2 / · · · À / En = E,
bEE } ÀA } AA w
EE }} AA
}}
} AA ww
EE
EE }}} AA
AA }} AA www
~} } A w
} }
~ w{ w
[1] [1] [1]
A1 A2 An
such that each Ai is isomorphic to up to shift a semistable sheaf on C. This

observation leads to the following Bridgeland’s stability conditions.
5.2. Bridgeland’s stability conditions

Here we introduce the notion of stability conditions on triangulated cate-
gories [14].
Definition 5.2. A stability condition on a triangulated category D consists of

data σ = (Z, P), where Z : K(D) → C is a linear map, and P(φ) ⊂ D is a full
additive subcategory for each φ ∈ R, which satisfy the following:
r P(φ + 1) = P(φ)[1].
r If φ1 > φ2 and Ai ∈ P(φi ), then Hom(A1 , A2 ) = 0.
r If E ∈ P(φ) is non-zero, then Z(E) = m(E) exp(iπφ) for some m(E) ∈
R>0 .
r For a non-zero object E ∈ D, we have the following collection of triangles:
0 = E0 / E1 / E2 / · · · À / En = E
bEE
EE }} ÀAA }} AA
AA x x
EE } AA } xx
EE }}} AA
A }}} AA
A xx
}
~ }
~ x| x
[1] [1] [1]
A1 A2 An
such that Aj ∈ P(φj ) with φ1 > φ2 > · · · > φn .
Here Z is called the central charge. Each P(φ) is an abelian category, the
non-zero objects of P(φ) are called semistable of phase φ, and the non-zero
simple objects of P(φ) are called stable. The objects Aj are called semistable
factors of E with respect to σ . We write φσ+ (E) = φ1 and φσ− (E) = φn . The
mass of E is defined to be

n
mσ (E) = |Z(Ai )|.
i=1
The following proposition is useful in constructing stability conditions.
Proposition 5.3. [14, Proposition 4.2] Giving a stability condition on D is

equivalent to giving the heart of a bounded t-structure A ⊂ D, and a group
homomorphism Z : K(D) → C called the stability function, such that for a
non-zero object E ∈ A one has
Z(E) ∈ {r exp(iπφ) | r > 0, 0 < φ ≤ 1},
and the pair (Z, A) satisfies the Harder-Narasimhan property.
The correspondence of Proposition 5.3 is obtained as follows. For an interval

I ⊂ R, let P(I ) ⊂ D be the minimal extension closed subcategory of D which
contains P(φ) with φ ∈ I . By Definition 5.2, one can easily check that the
subcategory P((0, 1]) ⊂ D is the heart of a bounded t-structure. Then,
(Z, P)
−→ (Z, P((0, 1])),
gives the correspondence of Proposition 5.3.

432 Yukinobu Toda
Remark 5.4. In general, the category P((a, b)) for b − a < 1 is not an abelian
category, and it is only a quasi-abelian category. (See [14, Section 4].)
Now we give some examples using Proposition 5.3.
Example 5.5. (i) Let C be a smooth projective curve and D = D(C). Then
setting A = Coh(C) and Z : K(C) → C by
Z(E) = − deg(E) + i rank(E),
gives a stability condition on D. In the language of Definition 5.2, P(φ) ⊂ D
for 0 < φ ≤ 1 is given by
P(φ) = {semistable sheaves E on C with μ(E) = −1/ tan(π φ).}
(ii) For a field k, let A be a finite dimensional k-algebra. Let modf (A)
be the abelian category of finite dimensional right A-modules, and D the
bounded derived category of modf (A). The abelian category modf (A) has
finite number of simple objects, say S1 , · · · , SN . Then setting A = modf (A)
and Z : K(D) → C arbitrary so that Im Z(Si ) > 0, (it is possible because the
classes [S1 ], · · · , [SN ] in K(D) form a basis of K(D)), gives a stability condi-
tion on D.
5.3. The space of stability conditions

For a stability condition σ = (Z, P) we say σ is locally finite if for any φ ∈ R
there is ε > 0 such that each quasi-abelian category P((φ − ε, φ + ε)) is of
finite length [14, Definition 5.7]. The set of locally finite stability conditions
on D is denoted by Stab(D). It has a natural topology induced by the metric,

mσ (E)
d(σ1 , σ2 ) = sup |φσ−2 (E) − φσ−1 (E)|, |φσ+2 (E) − φσ+1 (E)|, | log 2 |
0 =E∈D mσ1 (E)
∈ [0, ∞]. (5)
Forgetting the information of P, we have the map
Z : Stab(D) (Z, P)
−→ Z ∈ HomZ (K(D), C).
Theorem 5.6. [14, Theorem 1.2] For each connected component ⊂
Stab(D), there exists a linear subspace V () ⊂ HomZ (K(D), C), such that
Z restricts to a local homeomorphism, Z : → V ().
Remark 5.7. By the metric given by (5), we can easily see the following: for
a fixed E ∈ D, the functions
φ∗± (E), m∗ (E) : Stab(X) −→ R,
are continuous. In particular the set of points in Stab(D) where E is semistable

is closed in Stab(D).
In general Stab(D) is infinite dimensional, so we usually consider only
numerical stability conditions as in [14], [15].
Definition 5.8. Let X be a smooth projective variety and D = D(X). We say
a stability condition σ = (Z, P) on D is numerical if Z factors through the
Chern character map,
K(X)
Z / C.
uu:
u
uu
ch
uuu
u
H ∗ (X, Q)
We write Stab(X) for the set of locally finite numerical stability conditions on
D.
Let K0 (X) be the kernel of the map ch : K(X) → H ∗ (X, Q), and set N (X)
as follows:
N (X) = K(X)/K0 (X).
The group N (X) is called numerical Grothendieck group. Note that we have
the following injection,
1
ch : N (X) −→ H i (X, Z),
i!
hence in particular N (X) is a finitely generated Z-module. We have the fol-
lowing map,
Z : Stab(X) −→ Hom(N (X), C).
Because the dimension of Hom(N (X), C) is finite, Theorem 5.6 implies that
any connected component of Stab(X) is a finite dimensional complex manifold.
5.4. Group actions

Let GL+ (2, R) ⊂ GL(2, R) be the subgroup of linear maps T : R2 → R2
: + (2, R) → GL+ (2, R) be the uni-
which preserves orientation of R2 . Let GL
: + (2, R) is written as a pair (T , f ),
versal cover. Note that an element of GL
where f : R → R is an increasing map with f (φ + 1) = f (φ) + 1 and
T ∈ GL+ (2, R) such that induced maps on
S 1 = R/2Z = (R2 \ {0})/R>0 ,
are the same maps.
434 Yukinobu Toda
Lemma 5.9. For a triangulated category D, the space Stab(D) carries the
: + (2, R), and the left action of the group Auteq(D).
right action of the group GL
+
: (2, R), define
Proof. Given a stability condition σ = (Z, P) and (T , f ) ∈ GL

the stability condition σ = (Z , P ) by setting
Z = T −1 ◦ Z, P (φ) = P(f (φ)).
: + (2, R) on Stab(D).
Then it gives the right action of GL
Next for σ = (Z, P) ∈ Stab(D) and ∈ Auteq(D), its left action is defined
by (σ ) = (Z , P ) where
Z = Z ◦ −1
, P (φ) = (P(φ)).
+
: (2, R) and the restriction of the action
Note that there is a subgroup C ⊂ GL
: + (2, R) to C is free. Explicitly for λ ∈ C, its action on σ = (Z, P) is
of GL
described by σ = (Z , P ) where
Z = e−iπλ Z, P (φ) = P(φ + Re(λ)).
5.5. Background from string theory

The idea of the definition of the stability conditions comes from Douglas’s
'-stability for BPS-branes [22], [23]. Here we give the brief explanation for
the motivation of Definition 5.2. As we see later, the background from string
theory gives a nice explanation of the descriptions of the space of stability
conditions. However the author is not an expert in this area, and also there are
excellent explanations of the background from string theory for mathematicians
by Bridgeland himself [14, (1,4)], [9]. Thus for the detail and accuracies, the
author recommends consulting the articles [14], [9], [22], [23].
First let us consider a triple,
(X, β + iω, I ), (6)
where X is a simply connected Calabi-Yau 3-fold, I is a complex structure on

X, and β + iω ∈ H 2 (X, C) is a complexified Kähler class, i.e. ω is a Kähler
class. A triple (6) is called physicist’s Calabi-Yau 3-fold. Physicists believe that
for a triple (6), there is an associated non-linear sigma model, which defines
N = 2 super conformal field theory, simply referred to as SCFT. Here we note
that the sigma models are not at all defined mathematically. At least there is
an issue of convergence of the integrals in defining them, however we ignore
this problem. Let M be the the moduli space of SCFT. Then for a fixed X, the
subset UX ⊂ M which comes from sigma model constructions from the triples
(6) is an open subset. We call UX the neighborhood of the large volume limit,
and let M(X) be the connected component of M which contains UX . Here
we note that there might be some points in M(X) which correspond to sigma
models for topologically distinct Calabi-Yau 3-folds, or do not correspond to
sigma models at all.
There are two foliations on M, one of them is when restricted to UX , fixing
the complex structure I and deforming β + iω. Another foliation is fixing
β + iω and deforming I on UX . Let us consider the leaves of the corresponding
foliations,
MK (X) ⊂ M(X), MC (X) ⊂ M(X).
The space MC (X) is nothing but the moduli space of the complex structures,
which is defined mathematically and has been studied in algebraic geom-
etry up to now. However the space MK (X), called stringy Kähler moduli
space, is not defined mathematically rigorous way, as well as sigma model
constructions.
Now let us discuss on mirror symmetry. There is an involution m : M → M
called the mirror map, which exchanges the above two foliations on M. Then
a triple
(X̂, β̂ + i ω̂, Iˆ)
is called a mirror of the triple (6) if the associated sigma model is mapped to the
sigma model of (6) by the mirror map m. Especially we have the isomorphisms,
MK (X) ∼
= MC (X̂), MK (X̂) ∼
= MC (X).
Hence if we assume the mirror symmetry, it might be possible to define the

space MK (X) to be MC (X̂) for a mirror manifold X̂. However it is unsat-
isfactory, since it assumes the mirror symmetry, and it is not intrinsic. One
of the motivation of introducing the space of stability conditions is to define
MK (X) without using mirror symmetry, purely from the categorical data of
D(X). Actually it is expected that the space Stab(X) is related to the space
MK (X).
Next let us consider the D-branes. For a given SCFT, we have the two
associated topological conformal field theories (TCFT), called A-model and
B-model. If the SCFT is given by the triple (6), then the category of D-branes
of A and B-models are supposed to be
D b Fuk(X, β + iω), D b Coh(X, I ),

436 Yukinobu Toda
respectively. Here D b Coh(X, I ) is the bounded derived category of coher-

ent sheaves of the complex manifold (X, I ), and D b Fuk(X, β + iω) is the
derived Fukaya category, whose objects are (roughly speaking) local systems
on Lagrangian submanifolds on X. Note that D b Coh(X, I ) does not depend
on β + iω, and D b Fuk(X, β + iω) does not depend on I . The mirror map m
exchanges A and B models, hence there should exist equivalences,
D b Fuk(X, β + iω) ∼
= D b Coh(X̂, Iˆ), D b Coh(X, I ) ∼
= D b Fuk(X̂, β̂ + i ω̂),
for a mirror (X̂, β̂ + i ω̂, Iˆ). This is the Homological mirror symmetry proposed
by Kontsevich [45].
The fact that D b Coh(X, I ) does not depend on β + iω imply that the associ-
ated categories of B-models are unchanged along the subspace MK (X) ⊂ M.
Below we fix the complex structure I and simply write X for the complex man-
ifold (X, I ). The important point is that the category of particular B-branes,
called BPS-branes, are changed along MK (X) in the fixed category of B-
branes D(X). Let us take a point p ∈ MK (X) and the associated category of
BPS branes B(p). Here we remark that the point p does not have the informa-
tion how the category B(p) is embedded into D(X). Now let us take a closed
path,
γ : [0, 1] −→ MK (X),
with γ (0) = γ (1) = p, and choose an embedding B(p) ⊂ D(X). Then for each
t ∈ [0, 1], we have the associated subcategory,
B(γ (t)) ⊂ D(X).
For t = 1, we obtain the embedding of B(γ (1)) ⊂ D(X), and B(γ (1)) = B(p).
So we obtain the two embeddings of B(p) into D(X) for t = 0, 1, which
may be different. In fact its difference should be induced by the action of an
autoequivalence on D(X), hence we have the map,
π1 MK (X) −→ Auteq D(X).
Now recall that the space Stab(X) carries a left action of Auteq D(X). Also
it is expected that for σ = (Z, P) ∈ Stab(X), the associated category

P(φ),
φ∈R
defines the category of BPS branes at some point of MK (X). Combined these
observations, it is reasonable to guess that the space Auteq D(X)\ Stab(X) is
related to the space MK (X). More precisely Bridgeland conjectures in [9] that
the double quotient space
Auteq D(X)\ Stab(X)/C, (7)
contains the space MK (X). Hence in the viewpoint of string theory, it is

interesting to describe the space (7) and compare it with MC (X̂) of a mirror
manifold X̂.
Since Bridgeland wrote the paper [14], a number of papers on stability
conditions have appeared and some examples, properties have been studied.
Now at least there are following articles, [15], [10], [13], [59], [51], [47], [28],
[5], [63], [65]. Below we introduce some of them.
5.6. Stability conditions on elliptic curves

Let X be an elliptic curve. Then it is proved in [14] that the space Stab(X) is
: + (2, R) is free and transitive. Hence we have
connected and the action of GL
+
Stab(X) ∼ : (2, R),
= GL
: + (2, R) is induced
as a complex manifold. Here the complex structure on GL
by the local homeomorphism,
: + (2, R) → GL+ (2, R) ⊂ C2 .
GL
The group of autoequivalence is completely known by Theorem 2.11. Hence

one can describe the space (7), and the result is
Auteq D(X)\ Stab(X)/C ∼

= H/ SL(2, Z). (8)
Note that the RHS of (8) is nothing but the moduli space of elliptic curves.
Hence in this case, we have obtained the complete picture.
5.7. Stability conditions on K3 surfaces

Suppose X is a K3 surface. In this case, one of the connected components
Stab∗ (X) ⊂ Stab(X) is described by [15].
One of the difficult points in studying Stab(X) in this case is to show the
existence of stability conditions. Before discussing this, let us recall the notion
of μω -semistable sheaves. For a torsion free sheaf E ∈ Coh(X) and an ample
divisor ω on X, set μω (E) as
c1 (E) · ω
μω (E) = .
rank(E)
438 Yukinobu Toda
Definition 5.10. A torsion free sheaf E ∈ Coh(X) is called μω -semistable if

for any non-zero subsheaf F ⊂ E, one has
μω (F ) ≤ μω (E).
Remark 5.11. The notion of μω -semistability does not give the stability con-
ditions on D(X) directly. One may try to construct a stability condition as the
pair (Z, Coh(X)), where Z : K(X) → C is
Z(E) = −c1 (E) · ω + rank(E)i.
However it is not a stability condition because Z(Ox ) = 0 for closed points
x ∈ X.
Bridgeland [15] constructed stability conditions using the method of tilting.
Let β, ω be Q-divisors on X with ω ample. For a torsion free sheaf E ∈ Coh(X),
one has the Harder-Narasimhan filtration
0 = E0 ⊂ E1 ⊂ · · · ⊂ En−1 ⊂ En = E,
such that Fi = Ei /Ei−1 is μω -semistable and μω (Fi ) > μω (Fi+1 ). Then define
T(β,ω) ⊂ Coh(X) to be the subcategory consisting of sheaves whose torsion free
parts have μω -semistable Harder-Narasimhan factors of slope μω (Fi ) > β · ω.
Also define F(β,ω) ⊂ Coh(X) to be the subcategory consisting of torsion free
sheaves whose μω -semistable factors have slope μω (Fi ) ≤ β · ω.
Definition 5.12. We define A(β,ω) to be
A(β,ω) = {E ∈ D(X) | H−1 (E) ∈ F(β,ω) , H0 (E) ∈ T(β,ω) }.
The category A(β,ω) is called tilting with respect to the torsion pair
(T(β,ω) , F(β,ω) ). The next step is to construct the central charge Z(β,ω) : N (X) →
C. Let NS∗ (X) be the Mukai lattice,
NS∗ (X) = Z ⊕ NS(X) ⊕ Z.
For vi = (ri , li , si ) ∈ NS∗ (X) with i = 1, 2, its bilinear pairing is given by
(v1 , v2 ) = l1 · l2 − r1 s2 − r2 s1 . (9)
For an object E ∈ D(X) its Mukai vector is defined by,
8
v(E) = ch(E) tdX
= (r(E), c1 (E), ch2 (E) + r(E)).
Sending an object to its Mukai vector gives an isomorphism,
∼
=
v : N (X) −→ NS∗ (X). (10)
Under the identification (10), the bilinear pairing −χ (E1 , E2 ) on the left hand
side goes to the pairing (9). Let us define Z(β,ω) : N (X) → C by the formula.
Z(β,ω) (E) = (exp(β + iω), v(E)). (11)
Suppose v(E) = (r, l, s) with r = 0. Then (11) is written as
1 2
Z(β,ω) (E) = (l − 2rs) + r 2 ω2 − (l − rβ)2 + i(l − rβ) · ω. (12)
2r
If r = 0, (11) is written as Z(E) = (−s + l · β) + i(l · ω). We define σ(β,ω) to
be the pair (Z(β,ω) , A(β,ω) ).
Proposition 5.13. [15, Lemma 6.2, Proposition 7.1] The subcategory
A(β,ω) ⊂ D(X) is a heart of a bounded t-structure, and the pair σ(β,ω) gives a
stability condition on D(X) if and only if for any spherical sheaf E on X, one
has Z(β,ω) (E) ∈
/ R≤0 . This holds whenever ω2 > 2.
Let Stab∗ (X) be the connected component of Stab(X) which contains σ(β,ω) .
Note that the pairing (9) and the isomorphism (10) induces an isomorphism,
Hom(N (X), C) ∼
= NS∗ (X).
Hence we have the map,
Z : Stab∗ (X) −→ NS∗ (X).
In order to understand the image of Z, first define the subset
P(X) ⊂ NS∗ (X)C ,
to be the set of v = v1 + iv2 ∈ NS∗ (X)C with vi ∈ NS∗ (X) such that v1 , v2
span a positive definite two plane in NS∗ (X). The space P(X) consists of
two connected components, and let P + (X) ⊂ P(X) be the component which
contains (1, iω, − 12 ω2 ). We define P0+ (X) to be

P0+ (X) = P0 (X) \ δ⊥.
δ∈(X)
∗
Here (X) = {v ∈ NS (X) | (v, v) = −2}.
Theorem 5.14. (Bridgeland [15, Theorem 1.1]) The component Stab∗ (X) is
mapped by Z onto the open subset P0+ (X), and the induced map
Z : Stab∗ (X) −→ P0+ (X),
is a regular covering map. Its Galois group G fits into the exact sequence,
0 −→ G −→ Auteq∗ D(X) −→ Aut H ∗ (X, Z).
440 Yukinobu Toda
Here Auteq∗ D(X) is the group of autoequivalences of D(X) which preserve

the component Stab∗ (X).
Based on Theorem 5.14, Bridgeland proposes the following conjecture,
which relates the topology of Stab(X) and the group of autoequivalences of
D(X).
Conjecture 5.15. ([15, Conjecture 1.2]) The action of Auteq D(X) on Stab(X)
preserves the component Stab∗ (X). Furthermore Stab∗ (X) is simply connected.
As a consequence, we have the following short exact sequence,
1 −→ π1 P0+ (X) −→ Auteq D(X) −→ Aut+ H ∗ (X, Z) −→ 1.
Here Aut+ H ∗ (X, Z) is the index two subgroup of Aut H ∗ (X, Z) consisting
elements whose actions on P(X) preserve the component P + (X).
5.8. Stability conditions and birational geometry

Before giving another example, we point out the relationship between the
descriptions of stability conditions and birational geometry. For instance sup-
pose X is a minimal 3-fold. Recall that the birational minimal models are not
unique, but connected by a finite number of flops. The philosophy of Y. Kawa-
mata [36] is that one can capture the set of birational minimal models via a
chamber structures on the movable cone [36, Definition 1.1]. According to [36,
Theorem 2.3], chambers on the movable cone are given by the ample cones of
birational minimal models.
On the other hand, let us consider the space Stab(X). As explained in (5.5),
Stab(X) is related to the stringy Kähler moduli space MK (X). By restricting
the neighborhood of the large volume limit to MK (X), we obtain the open sub-
set UK (X) ⊂ MK (X). However there might be another topologically distinct
manifold W such that the associated neighborhood of the large volume limit
determines an open subset UK (W ) ⊂ MK (X). In this case, because the asso-
ciated TCFT of B-models are invariant along MK (X), we have an equivalence
of derived categories,
D(W ) −→ D(X),
i.e. W is a Fourier-Mukai partner of X. Recall that if W is birationally equivalent
to X, then W is a Fourier-Mukai partner of X. (See Theorem 3.12.) Thus in
describing MK (X), one has to take account the neighborhoods of the large
volume limits corresponding to several Fourier-Mukai partners such as flops.
In fact suppose X† → Y ← X is a flop at a single rational curve C ⊂ X. In
this case, P.Aspinwall [1, Figure 2] describes the localized picture of MK (X),
assuming the volumes of all the curves in X except C are big enough. The
resulting picture is
MK (X) = P1 \ {0, 1, ∞} (13)
†
= U K (X) ∪ U K (X ), (14)
and UK (X) ∩ UK (X† ) = ∅. Here string theory has singularities at one of the
deleted points {0, 1, ∞}, and the other two points are large volume limits
corresponding to X and X† .
Let us translate the above argument to the space Stab(X). First we can
guess that there exists some open subset UX ⊂ Stab(X), which corresponds to
the neighborhood of the large volume limit. For any point σ ∈ UX , the corre-
sponding t-structures given in Proposition 5.3 should be related to the standard
t-structure, for example obtained by tilting. Next let W and X be three dimen-
sional birational minimal models, and : D(W ) → D(X) an equivalence of
derived categories. Then induces an homeomorphism,
∗: Stab(W ) −→ Stab(X). (15)

Transferring UW by ∗ , we obtain a region ∗ UW ⊂ Stab(X). Conjecturally
there should exist a chamber structure of the following type,

∗ U W ⊂ Stab(X), (16)
(W, )
such that two chambers are either equal or disjoint. This picture is quite similar
to the movable cone [36, Theorem 2.3]. Thus by describing the space Stab(X),
we can observe the relationship between SCFT and birational geometry.
5.9. Stability conditions and 3-fold flops

In order to realize the picture (16), we give the description of the space of
stability conditions on a certain local Calabi-Yau 3-fold. Let f : X → Y be a
crepant small resolution of the conifold singularity,
f : X −→ Y = Spec C[x, y, z, w]/(xy − zw),
and g : X† → Y be its flop as in Example 3.6. Let C ∼
= P1 be the exceptional
locus of f . Set D(X/Y ) as follows,
D(X/Y ) = {E ∈ D(X) | Supp(E) ⊂ C}.
Let us describe the space of stability conditions on D(X/Y ), denoted by
Stab(X/Y ). As we will see, the description of Stab(X/Y ) realizes the pictures
(13), (14), (16), and also gives a nice picture for non-commutative crepant
442 Yukinobu Toda
resolutions discussed in Section 4. The situation we treat here is the simplest

case, and one can consult [63], [12], [65] for more general cases.
Let Coh(X/Y ), K(X/Y ) be
Coh(X/Y ) = D(X/Y ) ∩ Coh(X), K(X/Y ) = K(D(X/Y )).
Also denote by N 1 (X/Y ) the R-vector space of the numerical classes of R-

divisors, and A(X/Y )C ⊂ N 1 (X/Y )C the complexified ample cone,
A(X/Y )C = {β + iω ∈ N 1 (X/Y )C | ω is ample}.
In our case, we have
K(X/Y ) = Z[Ox ] ⊕ Z[OC (−1)], N 1 (X/Y )C = C[c1 (OX (1))] ∼

= C,
∼ {β + iω ∈ C | ω > 0}.
A(X/Y )C =
For β + iω ∈ N 1 (X/Y )C , set Z(β,ω) : K(X/Y ) → C as follows,
Z(β,ω) (E) = − ch3 (E) + (β + iω) · ch2 (E).
Lemma 5.16. ([63, Lemma 4.1]) For β + iω ∈ A(X/Y )C , the pair
σ(β,ω) = (Z(β,ω) , Coh(X/Y ))
determines a point of Stab(X/Y ). The set of points
UX = {σ(β,ω) ∈ Stab(X/Y ) | β + iω ∈ A1 (X/Y )C },
forms an open subset in the subspace Stabn (X/Y ),
Stabn (X/Y ) = {(Z, P) ∈ Stab(X/Y ) | Z([Ox ]) = −1},
which we call the set of normalized stability conditions.
It is easy to see that for Z : K(X/Y ) → C, we have Z([Ox ]) = −1 if and

only if Z is written as Z(β,ω) for some β + iω ∈ N 1 (X/Y )C . Hence we have
the following cartesian square,
Stabn (X/Y ) / Stab(X/Y )
Zn Z

N 1 (X/Y )C / K(X/Y )∗C .
The map Zn restricts to the homeomorphism between UX and A(X/Y )C .

Next let us see the behavior of the boundary of UX . Note that for each k ∈ Z,
the object OC (k − 1) is stable in any σ ∈ UX . Therefore by Remark 5.7, it
must be semistable in σ for σ ∈ ∂UX . Hence if we write σ = (Z(β,0) , P) with

β ∈ N 1 (X/Y ), we must have
Z(β,0) (OC (k − 1)) = −k + β = 0.
This implies
Zn (∂UX ) ⊂ R \ Z.
Conversely take β ∈ (k − 1, k). Then using the descriptions of the simple
objects given in Example 4.12, we can see that the pair

(Z(β,0) , 0 Per(X/Y ) ⊗ OX (k) ∩ D(X/Y )),
gives a stability condition and contained in ∂UX . Hence we have
Zn (∂UX ) = R \ Z.
We set ∂UX (k) to be
∂UX (k) = {σ ∈ ∂UX | Zn (σ ) ∈ (k − 1, k)}.
The next step is to consider the flop X† → Y . By Theorem 3.12, there are
equivalences,
: D(X† ) −→ D(X),
†
: D(X) −→ D(X† ),
such that takes −1 Per(X† /Y ) to 0 Per(X/Y ), and takes −1 Per(X/Y ) to
0 †
Per(X /Y ). The equivalence induces the following commutative diagram,
Stabn (X† /Y )
∗
/ Stabn (X/Y )

N 1 (X† /Y )C
φ∗
/ N 1 (X/Y )C .
Here φ : X† X is the birational map. By Theorem 3.12 and Example 4.12,
the equivalence restricts to the equivalence,
0
Per(X† /Y ) ⊗ OX† (1) ∩ D(X/Y ) −→ 0 Per(X/Y ) ∩ D(X/Y ).
This means ∗ induces a homeomorphism,
∗: ∂UX† (1) −→ ∂UX (0).

Also ⊗OX (k) induces a homeomorphism,
(⊗OX (k))∗ : ∂UX (0) −→ ∂UX (k).
444 Yukinobu Toda
Hence the two regions,

(⊗OX (k) ◦ )∗ UX† , UX ⊂ Stabn (X/Y )
are patched together in the boundary ∂UX (k). Repeating this argument, we
obtain the following theorem.
Theorem 5.17. ([63, Theorem 1.1, 1.2], [65, Theorem 7.3]) One has the
chamber structure,

Stab∗n (X/Y ) = (∗ U W ,
where Stab∗n (X/Y ) is the connected component of Stabn (X/Y ) which contains
UX , W is either X or X† , and ( : D(W ) → D(X) is written as a composition
of , † and tensoring line bundles OX (k), OX† (k ), i.e. written as
( = ⊗OX (k) ◦ ◦ ⊗OX† (k ) ◦ †
◦ ⊗OX (k ) · · · .
Furthermore the map Zn induces the map,
Stab∗n (X/Y ) −→ C \ Z,
which is a regular covering map. Its Galois group G is generated by the
spherical twist TOC (−1) and fits into the exact sequence,
1 −→ G −→ Auteq D(X/Y ) −→ Pic(X) × Z −→ 1,
where Auteq D(X/Y ) is the group of autoequivalences on D(X/Y ) which are
of Fourier-Mukai type with kernel supported on X ×Y X. Moreover we have
the homeomorphism,
Stab∗n (X/Y ) × C ∼
= Stab(X/Y ).
As a corollary, we have the following description of the double quotient (7),
Auteq D(X/Y )\ Stab(X/Y )/C ∼
=< TOC (−1) , ⊗OX (1) > \ Stab∗n (X/Y )
∼
= Z\(C \ Z)
∼
= P1 \ {0, 1, ∞}.
Hence we have obtained the same picture from the physics [1, Figure 2]. We
can divide the above double quotient into three pieces,
P1 \ {0, 1, ∞} = UX ∪ ∂UX ∪ UX† .
We see that UX , UX† correspond to the neighborhoods of the large volume limits
at X, X† respectively. The region ∂UX corresponds to the t-structure 0 Per(X/Y ),
which is according to Example 4.12, equivalent to a module category of a certain
non-commutative algebra. Hence we have realized the pictures (13), (14), and
(16).
5.10. Holomorphic generating functions on the

space of stability conditions
Finally, we introduce the recent work of D. Joyce [33] on the generating func-
tions on the space of stability conditions. Joyce’s attempt is to construct the
interesting holomorphic functions on Stab(D), using “counting invariants” of
semistable objects. This idea is motivated by the following fact. For a com-
pact symplectic manifold (M, ω), one can define the Gromov-Witten invariants
which count stable maps from Riemann surfaces, and the holomorphic gener-
ating function,
f : H ev (M, C) −→ C,
called Gromov-Witten potential, given by a power series whose coefficients

are Gromov-Witten invariants. The function f satisfies a partial differential
equation (p.d.e), called WDVV equation, and it defines the Frobenius structure
on H ev (M, C).
The purpose of the paper [33] is to develop the similar theory on the space of
stability conditions. However for several technical reasons, his arguments work
only on the space of stability conditions on certain special abelian categories,
which includes the case of Example 5.5 (ii). Let A be a finite dimensional
k-algebra, and A = modf (A), D, and S1 , · · · , SN ∈ A be as in Example 5.5
(ii). The space Stab(D) has the following subset,
Stab(A) = {(Z, P) ∈ Stab(D) | P((0, 1]) = A and Im Z([Si ]) > 0 for all i}.
It is obvious Stab(A) is isomorphic to HN , where H = {a + bi ∈ C | b > 0}.

We put the following additional assumption.
Assumption 5.18. There is a bilinear form χ : K(A) × K(A) → Z, such that

for E, F ∈ A, we have
χ(E, F ) = dim Hom(E, F ) − dim Ext1 (E, F ) + dim Ext1 (F, E)

− dim Hom(F, E). (17)
Here we give some situations in which the above assumption is satisfied.
Example 5.19. (i) Suppose for E, F ∈ A one has dim Exti (E, F ) = 0 for
i ≥ 2. Then
χ(E, F ) = χ (E, F ) − χ (F, E),
satisfies (17).
446 Yukinobu Toda
(ii) Suppose A is a three dimensional Calabi-Yau algebra, i.e. E

→ E[3]
gives a Serre functor on D. Then the usual Euler pairing

3
χ (E, F ) = (−1)i dim Exti (E, F ),
i=0
satisfies (17) by Serre duality. So we can set χ = χ .
Using χ, we construct the Lie algebra over C,

L= Lα , Lα = C · cα ,
α∈K(A)
with
[cα , cβ ] = χ (α, β)cα+β .
Jacobi identity follows because χ is anti-symmetric by (17).

We set C(A) ⊂ K(A) as follows,
C(A) := Im(A → K(A)) \ {0},
and take α ∈ C(A). Then Joyce constructs the holomorphic function
f α : Stab(A) −→ C,
using counting invariants of semistable objects and satisfies certain p.d.e. Now
we give the outline of the construction of f α in [33]. For each α ∈ C(A) and
σ ∈ Stab(A), Joyce [34] associates the invariant, J α (σ ) ∈ Q, as a counting
invariant of semistable objects E with [E] = α. The definition of J α (σ ) is
similar to the virtual Euler number of the moduli space of σ -semistable objects
of type α. It is known that there is a nice moduli space of semistable A-
modules, so its virtual Euler number is well-defined. (See [41].) However for
our purpose, it doesn’t work well, when there is a σ -semistable object which
is not stable. Actually we have to take account of the automorphism groups of
semistable objects, so Joyce works in the context of moduli stacks of semistable
objects, which are known to be algebraic stacks. Roughly speaking J α (σ ) is a
“weighted Euler number” of the moduli stack of σ -semistable objects of type
α, and especially when α is primitive and σ is generic, the invariant J α (σ )
coincides with the virtual Euler number of its coarse moduli space. For the
detail, see [34]. We set
α (σ ) as follows,

α (σ ) = J α (σ )cα ∈ Lα .
Here we remark that the function σ

→
α (σ ) is discontinuous in general. In
fact there is a wall and chamber structure on Stab(A), such that the set of
semistable objects does not change in a chamber. (See [15].) This implies that
σ
→
α (σ ) is constant in a chamber, however it will be a different value under a
wall crossing. The idea of [33] is to cancel out the discontinuity of σ
→
α (σ ),
by using the (also discontinuous) functions,
Fn : (C∗ )n −→ C,
and make a function f α : Stab(A) → U (L),

f α (σ ) = Fn (Z(α1 ), · · · , Z(αn ))
α1 (σ ) ∗ · · · ∗
αn (σ ). (18)
α1 +···+αn =α
Here U (L) is the universal enveloping algebra of L, ∗ is the multiplication in

U (L), αi ∈ C(A), and σ = (Z, A) ∈ Stab(A). Note that the sum (18) is a finite
sum by our definition of A. (This is one of the reasons why the arguments
in [33] are restricted in this case.) The following is the main theorem of [33].
Theorem 5.20. (Joyce [33]) There is a family of functions Fn : (C∗ )n → C
such that
r f α is contained in Lv ∼
= C. Hence defines a function f α : Stab(A) → C.
r The function f α is continuous and holomorphic.
Furthermore Fn is uniquely determined (under several normalizations), with
F1 , F2 as follows,
. /
1 1
log zz21 − π i , zz21 ∈
/ (0, ∞),
F1 (z1 ) = , F2 (z1 , z2 ) = (2πi) 2
2π i 1
2 log
(2πi)
z2
, z1
z 2
∈ (0, ∞).
z1
Here the branch of log is determined by log(−1) = π i. The functions f α for

α ∈ C(A) satisfy the following p.d.e,
dZ(β)
df α (σ ) = −[f β (σ ), f γ (σ )] ⊗ ,
β+γ =α
Z(β)
where β, γ ∈ C(A).
Explicitly the function f α is expanded in the following formula,

n
f α (σ ) = Fn (Z(α1 ), · · · , Z(σn )) J αi (σ )
α1 +···+αn =α i=1
⎡ ⎤
1
×⎣ χ (αi , αj )⎦ ,
2n−1 i→j in
with αi ∈ C(A), is a connected and simply connected oriented graph with
vertex {1, · · · , n}, and if there is an arrow i → j in , we must have i ≤ j .
448 Yukinobu Toda
At this time extending Joyce’s work to triangulated categories is problematic,

even if we ignore the convergence of the sum. In the proof of Theorem 5.20,
he uses his works on configurations in abelian categories [30], [31], [32], [34],
so we have to generalize these works to triangulated categories. The author’s
work [62] on the invariants of semistable objects on K3 surfaces is toward one
step for this goal.
Acknowledgement
The author thanks Hokuto Uehara for checking the manuscript and giving the
nice advice. He is supported by Japan Society for the Promotion of Sciences
Research Fellowships for Young Scientists, No 198007.
References
[1] P. Aspinwall. A Point’s Point of View of Stringy Geometry. J. High Energy Phys,
Vol. 002, p. 15pp, 2003.
[2] A. Beilinson. Coherent sheaves on Pn and problems of linear algebra. Funct. Anal.
Appl, Vol. 12, pp. 214–216, 1978.
[3] A. Beilinson, J. Bernstein, and P. Deligne. Faisceaux pervers. Analysis and topol-
ogy on singular spaces I, Asterisque, Vol. 100, pp. 5–171, 1982.
[4] O. Ben-Bassat, J. Block, and T. Pantev. Non-commutative tori and Fourier-Mukai
duality. Compositio Math, Vol. 143, pp. 423–475, 2007.
[5] A. Bergman. Stability conditions and Branes at Singularities. preprint.
math.AG/0702092.
[6] R. Bezrukavnikov and D. Kaledin. McKay equivalence for symplectic resolutions
of quotient singularities. Proc. Steklov Inst. Math, Vol. 246, pp. 13–33, 2004.
[7] A. Bondal and D. Orlov. Semiorthgonal decomposition for algebraic varieties.
preprint. math.AG/9506012.
[8] A. Bondal and D. Orlov. Reconstruction of a variety from the derived category
and groups of autoequivalences. Compositio Math, Vol. 125, pp. 327–344, 2001.
[9] T. Bridgeland. Spaces of stability conditions. preprint. math.AG/0611510.
[10] T. Bridgeland. Stability conditions and Kleinian singularities. preprint.
math.AG/0508257.
[11] T. Bridgeland. Flops and derived categories. Invent. Math, Vol. 147, pp. 613–632,
2002.
[12] T. Bridgeland. Derived categories of coherent sheaves. Proceedings of the 2006
ICM, 2006. math.AG/0602129.
[13] T. Bridgeland. Stability conditions on a non-compact Calabi-Yau threefold.
Comm. Math. Phys., Vol. 266, pp. 715–733, 2006.
[14] T. Bridgeland. Stability conditions on triangulated categories. Ann. of Math,
Vol. 166, pp. 317–345, 2007.
[15] T. Bridgeland. Stability conditions on K3 surfaces. Duke Math. J., Vol. 141,
pp. 241–291, 2008.
[16] T. Bridgeland, A. King, and M. Reid. The McKay correspondence as an

equivalence of derived categories. J. Amer. Math. Soc., Vol. 14, pp. 535–554,
2001.
[17] T. Bridgeland and A. Maciocia. Complex surfaces with equivalent derived cate-
gories. Math. Z, Vol. 236, pp. 677–697, 2001.
[18] A. Cǎldǎraru. The Mukai pairing, I: The Hochschild structure. preprint.
math.AG/0308079.
[19] J-C. Chen. Flops and equivalences of derived categories for three-folds with only
Gorenstein singularities. J. Differential. Geom, Vol. 61, pp. 227–261, 2002.
[20] M. Van den Bergh. Non-commutative crepant resolutions. The legacy of Niels
Henrik Abel, Springer, Berlin, pp. 749–770, 2004.
[21] M. Van den Bergh. Three dimensional flops and noncommutative rings. Duke
Math. J., Vol. 122, pp. 423–455, 2004.
[22] M. Douglas. D-branes, categories and N = 1 supersymmetry. J. Math. Phys.,
Vol. 42, pp. 2818–2843, 2001.
[23] M. Douglas. Dirichlet branes, homological mirror symmetry, and stability. Pro-
ceedings of the 1998 ICM, pp. 395–408, 2002. math.AG/0207021.
[24] R. Hartshorne. Residues and Duality : Lecture Notes of a Seminar on the Work of
A.Grothendieck, given at Harvard 1963/1964, Lecture Notes in Mathematics, No.
20, Springer-Verlag, 1966.
[25] L. Hille and M. Van den Bergh. Fourier-Mukai tranforms. preprint.
math.AG/0402043.
[26] G. Hochschild, B. Kostant, and A. Rosenberg. Differential forms on regular affine
algebras. Trans. AMS, Vol. 102, pp. 383–408, 1962.
[27] P. Horja. Derived Category Automorphisms from Mirror Symmetry. Duke Math. J.,
Vol. 127, pp. 1–34, 2005.
[28] A. Ishii, K. Ueda, and H. Uehara. Stability Conditions on An -Singularities.
math.AG/0609551.
[29] A. Ishii and H. Uehara. Autoequivalences of derived categories on the minimal
resolutions of An -singularities on surfaces. J. Differential Geom., Vol. 71, pp. 385–
435, 2005.
[30] D. Joyce. Configurations in abelian categories I. Basic properties and moduli stack.
Advances in Math, Vol. 203, pp. 194–255, 2006.
[31] D. Joyce. Configurations in abelian categories II. Ringel-Hall algebras. Advances
in Math, Vol. 210, pp. 635–706, 2007.
[32] D. Joyce. Configurations in abelian categories III. Stability conditions and identi-
ties. Advances in Math, Vol. 215, pp. 153–219, 2007.
[33] D. Joyce. Holomorphic generating functions for invariants counting coherent
sheaves on Calabi-Yau 3-folds. Geometry and Topology, Vol. 11, pp. 667–725,
2007.
[34] D. Joyce. Configurations in abelian categories IV. Invariants and changing stability
conditions. Advances in Math, Vol. 217, pp. 125–204, 2008.
[35] Y. Kawamata. Flops connect minimal models. preprint. math.AG/0704.1013.
[36] Y. Kawamata. On the cone of divisors of Calabi-Yau fiber spaces. Internat. J. Math,
Vol. 5, pp. 665–687, 1997.
[37] Y. Kawamata. D-equivalence and K-equivalence. J. Differential Geom., Vol. 61,
pp. 147–171, 2002.
450 Yukinobu Toda
[38] Y. Kawamata. Francia’s flip and derived categories. Algebraic geometry, de

Gruyter, Berlin, pp. 197–215, 2002.
[39] Y. Kawamata. Log crepant birational maps and derived categories.
J. Math. Sci. Univ. Tokyo, Vol. 12, pp. 1–53, 2005.
[40] Y. Kawamata. Derived categories of toric varieties. Michigan Math. J., Vol. 54,
pp. 517–535, 2006.
[41] A. King. Moduli of representations of finite-dimensional algebras.
Quart. J. Math. Oxford Ser.(2), Vol. 45, pp. 515–530, 1994.
[42] J. Kollár. Flops. Nagoya Math. J, Vol. 113, pp. 15–36, 1989.
[43] J. Kollár. Rational curves on algebraic varieties, Vol. 32 of Ergebnisse Math. Gren-
zgeb.(3). Springer-Verlag, 1996.
[44] J. Kollár and S. Mori. Birational geometry of algebraic varieties, Vol. 134 of
Cambridge Tracts in Mathematics. Cambridge University Press, 1998.
[45] M. Kontsevich. Homological algebra of mirror symmetry, Vol. 1 of Proceedings
of ICM. Basel:Birkhäuser, 1995.
[46] W. Lowen and M. Van den Bergh. Deformation theory of abelian categories.
Trans. Amer. Math. Soc., pp. 5441–5483, 2006.
[47] E. Macri. Some examples of moduli spaces of stability conditions on derived
categories. preprint. math.AG/0411613.
[48] S. Mukai. Duality between D(X) and D(X̂) with its application to picard sheaves.
Nagoya Math. J., Vol. 81, pp. 101–116, 1981.
[49] S. Mukai. On the moduli space of bundles on K3 surfaces I . Vector Bundles on
Algebraic Varieties, M. F. Atiyah et al., Oxford University Press, pp. 341–413,
1987.
[50] Y. Namikawa. Mukai flops and derived categories. J. Reine. Angew. Math., pp. 65–
76, 2003.
[51] S. Okada. Stability manifold of P1 . J.Algebraic Geom, Vol. 15, pp. 487–505, 2006.
[52] D. Orlov. On Equivalences of derived categories and K3 surfaces. J. Math. Sci
(New York), Vol. 84, pp. 1361–1381, 1997.
[53] D. Orlov. Derived categories of coherent sheaves on abelian varieties and equiv-
alences between them. Izv. Ross. Akad. Nauk. Ser. Mat., Vol. 66, pp. 131–158,
2002.
[54] D. Orlov. Derived categories of coherent sheaves and motives.
Uspekhi. Math. Nauk, Vol. 60, pp. 231–232, 2005.
[55] M. Reid. McKay correspondence. preprint. math.AG/9702016.
[56] P. Seidel. Graded Lagrangian submanifolds. Bull. Soc. Math. France, Vol. 128,
pp. 103–149, 2000.
[57] P. Seidel and R. P. Thomas. Braid group actions on derived categories of coherent
sheaves. Duke Math. J., Vol. 108, pp. 37–107, 2001.
[58] R. G. Swan. Hochschild cohomology of quasiprojective schemes. J. Pure
Appl. Algebra, Vol. 1, pp. 57–80, 1996.
[59] R. P. Thomas. Stability conditions and the braid groups. Comm. Anal. Geom.,
Vol. 14, pp. 135–161, 2006.
[60] Y. Toda. Fourier-Mukai transforms and canonical divisors. Compositio Math, Vol.
142, pp. 962–982, 2006.
[61] Y. Toda. On a certain generalization of spherical twists. Bulletin de la SMF, Vol.
135, pp. 97–112, 2007.
[62] Y. Toda. Moduli stacks and invariants of semistable objects on K3 surfaces.

Advances in Math, Vol. 217, pp. 2736–2781, 2008.
[63] Y. Toda. Stability conditions and crepant small resolutions.
Trans. Amer. Math. Soc., Vol. 360, pp. 6149–6178, 2008.
[64] Y. Toda. Deformations and Fourier-Mukai transforms. J. Differential Geom.,
Vol. 81, pp. 197–224, 2009.
[65] Y. Toda. Stability conditions and Calabi-Yau fibrations. J. Algebraic Geom.,
Vol. 18, pp. 101–133, 2009.
[66] H. Uehara. An example of Fourier-Mukai partners of minimal elliptic surfaces.
Math. Res. Lett., Vol. 11, pp. 371–375, 2004.
[67] A. Yekutieli. The continuous Hochschild cochain complex of a scheme. Canada
J. Math., Vol. 6, pp. 1319–1337, 2002.
Yukinobu Toda, Graduate School of Mathematical Sciences, University of

Tokyo
E-mail address:toda@ms.u-tokyo.ac.jp
Rigid dualizing complexes via differential
graded algebras (survey)
amnon yekutieli
Abstract. In this article we survey recent results on rigid dualizing complexes
over commutative algebras. We begin by recalling what are dualizing complexes.
Next we define rigid complexes, and explain their functorial properties. Due to
the possible presence of torsion, we must use differential graded algebras in the
constructions. We then discuss rigid dualizing complexes. Finally we show how
rigid complexes can be used to understand Cohen-Macaulay homomorphisms
and relative dualizing sheaves.
0. Introduction
This short article is based on a lecture I gave at the “Workshop on Triangulated
Categories”, Leeds, August 2006. It is a survey of recent results on rigid
dualizing complexes over commutative rings. Most of these results are joint
work of mine with James Zhang. The idea of rigid dualizing complex is due to
Michel Van den Bergh.
By default all rings considered in this article are commutative. We begin by
recalling the notion of dualizing complex over a noetherian ring A. Next let B
be a noetherian A-algebra. We define what is a rigid complex of B-modules
relative to A. In making this definition we must use differential graded algebras
(when B is not flat over A). The functorial properties of rigid complexes are
explained. We then discuss rigid dualizing complexes, which by definition
are complexes that are both rigid and dualizing. Finally we show how rigid
complexes can be used to understand Cohen-Macaulay homomorphisms and
relative dualizing sheaves.
I wish to thank my collaborator James Zhang. Thanks also to Luchezar
Avramov, Srikanth Iyengar and Joseph Lipman for discussions regarding the
material in Section 5.
Key words and phrases. commutative rings, DG algebras, derived categories, rigid complexes.
Mathematics Subject Classification 2000. Primary: 18E30; Secondary: 18G10, 16E45, 18G15.
This research was supported by the US-Israel Binational Science Foundation.
452
Rigid dualizing complexes 453
1. Dualizing Complexes: Overview

Let A be a noetherian ring. Denote by Dbf (Mod A) the derived category
of bounded complexes of A-modules with finitely generated cohomology
modules.
Definition 1.1. (Grothendieck [RD]) A dualizing complex over A is a complex

R ∈ Dbf (Mod A) satisfying the two conditions:
(i) R has finite injective dimension.

(ii) The canonical morphism A → RHomA (R, R) is an isomorphism.
Condition (i) means that there is an integer d such that ExtiA (M, R) = 0 for
all i > d and all modules M.
Recall that a noetherian ring K is called regular if all its local rings Kp ,
p ∈ Spec K, are regular local rings.
Example 1.2. If K is a regular noetherian ring of finite Krull dimension (say

a field, or the ring of integers Z) then
R := K ∈ Dbf (Mod K)
is a dualizing complex over K.
Dualizing complexes over commutative rings are part of Grothendieck’s

duality theory in algebraic geometry, which was developed in [RD]. This duality
theory deals with dualizing complexes on schemes and relations between them.
See Remark 4.4.
In Section 4 we explain a new approach to dualizing complexes over com-
mutative rings, due to James Zhang and the author (see [YZ4] and [YZ5]).
Specifically, we discuss existence and uniqueness of rigid dualizing complexes.
A dualizing complex R has many automorphisms; indeed, its group of auto-
morphisms in D(Mod A) is the group A× of invertible elements. The purpose
of rigidity is to eliminate automorphisms, and to make dualizing complexes
functorial. See Theorem 4.2.
In a sequel paper [Ye2] we use the technique of perverse coherent sheaves
to construct rigid dualizing complexes on schemes, and we reproduce almost
all of the geometric Grothendieck duality theory.
Related work in noncommutative algebraic geometry (where rigid dualizing
complexes were first introduced) can be found in [VdB, YZ1, YZ2, YZ3].
454 Amnon Yekutieli
2. Rigid Complexes and DG Algebras

Let me start with a discussion of rigidity for algebras over a field. Suppose K
is a field, B is a K-algebra, and M ∈ D(Mod B).
According to Van den Bergh [VdB] a rigidifying isomorphism for M is an
isomorphism

ρ : M → RHomB⊗K B (B, M ⊗K M) (2.1)
in D(Mod B).
Now suppose A is any ring. Trying to write A instead of K in formula (2.1)
does not make sense: instead of M ⊗A M we must take the derived tensor
product M ⊗LA M; but then there is no obvious way to make M ⊗LA M into a
complex of B ⊗A B -modules.
The problem is torsion: B might fail to be a flat A-algebra. This is where
differential graded algebras (DG algebras) enter the picture.

A DG algebra is a graded ring Ã = i∈Z Ãi , together with a graded deriva-
tion d : Ã → Ã of degree 1, satisfying d ◦ d = 0.
A DG algebra quasi-isomorphism is a homomorphism f : Ã → B̃ respect-
ing degrees, multiplications and differentials, and such that H(f ) : HÃ → HB̃
is an isomorphism (of graded algebras).
We shall only consider super-commutative non-positive DG algebras. Super-
commutative means that ab = (−1)ij ba and c2 = 0 for all a ∈ Ãi , b ∈ Ãj and

c ∈ Ã2i+1 . Non-positive means that Ã = i≤0 Ãi .
We view a ring A as a DG algebra concentrated in degree 0. Given a DG
algebra homomorphism A → Ã we say that Ã is a DG A-algebra.
Let A be a ring. A semi-free DG A-algebra is a DG A-algebra Ã, such
that after forgetting the differential Ã is isomorphic, as graded A-algebra, to a
super-polynomial algebra on some graded set of variables.
Definition 2.2. Let A be a ring and B an A-algebra. A semi-free DG algebra

resolution of B relative to A is a quasi-isomorphism B̃ → B of DG A-algebras,
where B̃ is a semi-free DG A-algebra.
Such resolutions always exist, and they are unique up to quasi-isomorphism.
Example 2.3. Take A = Z and B = Z/(6). Define B̃ to be the super-

polynomial algebra A[ξ ] on the variable ξ of degree −1. So B̃ = A ⊕ Aξ as
free graded A-module, and ξ 2 = 0. Let d(ξ ) := 6. Then B̃ → B is a semi-free
DG algebra resolution of B relative to A.
For a DG algebra Ã one has the category DGMod Ã of DG Ã-modules. It

is analogous to the category of complexes of modules over a ring, and by a
similar process of inverting quasi-isomorphisms we obtain the derived category

D̃(DGMod Ã); see [Ke], [Hi].
For a ring A (i.e. a DG algebra concentrated in degree 0) we have
D̃(DGMod A) = D(Mod A),
the usual derived category.

It is possible to derive functors of DG modules, again in analogy to
D(Mod A). An added feature is that for a quasi-isomorphism Ã → B̃, the
restriction of scalars functor
D̃(DGMod B̃) → D̃(DGMod Ã)
is an equivalence.
Getting back to our original problem, suppose A is a ring and B is an A-
algebra. Choose a semi-free DG algebra resolution B̃ → B relative to A. For
M ∈ D(Mod B) define
SqB/A M := RHomB̃⊗A B̃ (B, M ⊗LA M)
in D(Mod B).
Theorem 2.4. ([YZ4]) The functor
SqB/A : D(Mod B) → D(Mod B)
is independent of the resolution B̃ → B.
The functor SqB/A , called the squaring operation, is nonlinear. In fact, given
a morphism φ : M → M in D(Mod B) and an element b ∈ B one has
SqB/A (bφ) = b2 SqB/A (φ) (2.5)
in
HomD(Mod B) (SqB/A M, SqB/A M).
Definition 2.6. Let B be a noetherian A-algebra, and let M be a complex in

Dbf (Mod B) that has finite flat dimension over A. Assume

ρ : M → SqB/A M
is an isomorphism in D(Mod B). Then the pair (M, ρ) is called a rigid complex
over B relative to A.
Definition 2.7. Say (M, ρ) and (N, σ ) are rigid complexes over B relative to
A. A morphism φ : M → N in D(Mod B) is called a rigid morphism relative
456 Amnon Yekutieli
to A if the diagram
ρ
M −−−−→ SqB/A M
⏐ ⏐
⏐
φ
⏐Sq (φ)
B/A
σ
N −−−−→ SqB/A N
is commutative.
We denote by Dbf (Mod B)rig/A the category of rigid complexes over B relative
to A.
Example 2.8. Take M = B = A. Then
SqA/A A = RHomA⊗A A (A, A ⊗A A) = A,
and we interpret this as a rigidifying isomorphism

ρ tau : A → SqA/A A.
The tautological rigid complex is
(A, ρ tau ) ∈ Dbf (Mod A)rig/A .
3. Properties of Rigid Complexes

The first property of rigid complexes explains their name.
Theorem 3.1. ([YZ4]) Let A be a ring, B a noetherian A-algebra, and
(M, ρ) ∈ Dbf (Mod B)rig/A .
Assume the canonical ring homomorphism
B → HomD(Mod B) (M, M)
is bijective. Then the only automorphism of (M, ρ) in Dbf (Mod B)rig/A is the
identity 1M .
The proof is very easy: an automorphism φ of M has to be of the form

φ = b 1M for some invertible element b ∈ B. If φ is rigid then b = b2 (cf.
formula (2.5)), and hence b = 1.
We find it convenient to denote ring homomorphisms by f ∗ etc. Thus a ring
homomorphism f ∗ : A → B corresponds to the morphism of schemes
f : Spec B → Spec A.
Let A be a noetherian ring. Recall that an A-algebra B is called essentially

finite type if it is a localization of some finitely generated A-algebra. We say
that B is essentially smooth (resp. essentially étale) over A if it is essentially
finite type and formally smooth (resp. formally étale).
Example 3.2. If A is a localization of A then A → A is essentially étale. If
B = A[t1 , . . . , tn ] is a polynomial algebra then A → B is smooth, and hence
also essentially smooth.
Let A be a noetherian ring and f ∗ : A → B an essentially smooth homo-
morphism. Then #1B/A is a finitely generated projective B-module. Let
&
Spec B = Spec Bi
i
be the decomposition into connected components, and for every i let ni be the
rank of #1Bi /A . We define a functor
f & : D(Mod A) → D(Mod B)
by

f & M := #nBii /A [ni ] ⊗A M.
i
Recall that a ring homomorphism f ∗ : A → B is called finite if B is a
finitely generated A-module. Given such a finite homomorphism we define a
functor
f : D(Mod A) → D(Mod B)
by
f M := RHomA (B, M).
Theorem 3.3. ([YZ4]) Let A be a noetherian ring, let B, C be essentially finite
type A-algebras, let f ∗ : B → C be an A-algebra homomorphism, and let
(M, ρ) ∈ Dbf (Mod B)rig/A .
(1) If f ∗ is finite and f M has finite flat dimension over A, then f M has an
induced rigidifying isomorphism

f (ρ) : f M → SqC/A f M.
The assignment

(M, ρ)
→ f (M, ρ) := f (ρ), f M
is functorial.
458 Amnon Yekutieli
(2) If f ∗ is essentially smooth then f & M has an induced rigidifying isomor-

phism

f & (ρ) : f & M → SqC/A f & M.
The assignment

(M, ρ)
→ f & (M, ρ) := f & (ρ), f & M
is functorial.
4. Rigid Dualizing Complexes

Let K be a regular noetherian ring of finite Krull dimension. We denote by
EFTAlg /K the category of essentially finite type K-algebras.
Definition 4.1. A rigid dualizing complex over A relative to K is a rigid

complex (RA , ρA ), such that RA is a dualizing complex.
Theorem 4.2. ([YZ5]) Let K be a regular finite dimensional noetherian ring,

and let A be an essentially finite type K-algebra.
(1) The algebra A has a rigid dualizing complex (RA , ρA ), which is unique up
to a unique rigid isomorphism.
(2) Given a finite homomorphism f ∗ : A → B, there is a unique rigid isomor-

phism f (RA , ρA ) → (RB , ρB ).
(3) Given an essentially smooth homomorphism f ∗ : A → B , there is a unique

rigid isomorphism f & (RA , ρA ) → (RB , ρB ).
Here is how the rigid dualizing complex (RA , ρA ) is obtained. We begin

with the tautological rigid complex
(K, ρ tau ) ∈ Dbf (Mod K)rig/K ,
which is dualizing (cf. Examples 1.2 and 2.8). Now the structural homomor-
phism K → A can be factored into
f∗ g∗ h∗
K−
→B−
→C−
→ A,
where f ∗ is essentially smooth (B is a polynomial algebra over K); g ∗ is finite

(a surjection); and h∗ is also essentially smooth (a localization).
It is not hard to check (see [RD, Chapter V]) that each of the complexes
f & K, g f & K and h& g f & K is dualizing over the respective ring. In particular,
g f & K has bounded cohomology, and hence it has finite flat dimension over K.
According to Theorem 3.3 we then have a rigid complex

(RA , ρA ) := h& g f & (K, ρ tau ) ∈ Dbf (Mod A)rig/K .
Definition 4.3. Given a homomorphism f ∗ : A → B in EFTAlg /K, define the
twisted inverse image functor
f ! : D+ +
f (Mod A) → Df (Mod B)
by the formula

f ! M := RHomB B ⊗LA RHomA (M, RA ), RB .
It is easy to show that the assignment f ∗
→ f ! is a pseudofunctor from
the category EFTAlg /K to the 2-category Cat of all categories. Moreover,
using Theorem 4.2 one can show that this operation has very good properties.
For instance, when f ∗ is finite, then there is a functorial nondegenerate trace
morphism
Trf : f ! M → M.
Remark 4.4. According to Grothendieck’s duality theory in [RD], if f : X →
Y is a finite type morphism between noetherian schemes, and if Y has a dualizing
complex, then there is a functor
f !(G) : D+ +
c (Mod OY ) → Dc (Mod OX ),
with many good properties.

Let FTAlg /K be the category of finite type K-algebras. By restricting
attention to affine schemes, the results of [RD] give rise to a pseudofunctor
f ∗
→ f !(G) from FTAlg /K to Cat. It is not hard to show that the pseudofunc-
tor f ∗
→ f !(G) is isomorphic to our 2-functor f ∗
→ f ! ; see [YZ5, Theorem
4.10].
It should be noted that our construction works in the slightly bigger category
EFTAlg /K. It also has the advantage of being local; whereas in [RD] some of
the results require that morphisms between affine schemes be compactified.
5. Rigid Complexes and CM Homomorphisms

In this final section we discuss the relation between rigid complexes and Cohen-
Macaulay homomorphisms.
Definition 5.1. A ring A is called tractable if there is an essentially finite
type homomorphism K → A, for some regular noetherian ring of finite Krull
dimension K.
460 Amnon Yekutieli
Such a homomorphism K → A is called a traction for A. It is not part of

the structure – the ring A does not come with any preferred traction. “Most
commutative noetherian rings we know” are tractable.
Given a traction K → A we denote by RA/K the rigid dualizing complex
of A relative to K; cf. Theorem 4.2. (The rigidifying isomorphism ρA/K is
implicit.)
Recall that a noetherian ring A is called Cohen-Macaulay (resp. Gorenstein)
if all its local rings Ap , p ∈ Spec A, are Cohen-Macaulay (resp. Gorenstein)
local rings. The implications are regular ⇒ Gorenstein ⇒ Cohen-Macaulay.
Let f ∗ : A → B be a ring homomorphism. For p ∈ Spec A let k(p) :=
(A/p)p , the residue field. The fiber of f ∗ above p is the k(p)-algebra B ⊗A k(p).
Now assume f ∗ is an essentially finite type flat homomorphism. If all the
fibers of f ∗ are Cohen-Macaulay (resp. Gorenstein) rings, then we call f ∗ an
essentially Cohen-Macaulay (resp. essentially Gorenstein) homomorphism.
Theorem 5.2. ([Ye2]) Let A be a tractable ring, and let f ∗ : A → B be

homomorphism which is of essentially finite type and of finite flat dimension.
Then there exists a rigid complex RB/A over B relative to A, unique up to a
unique rigid isomorphism, with the following property:
(*) Let K → A be some traction. Then
RA/K ⊗LA RB/A ∼

= RB/K
in D(Mod B).
Condition (*) implies that the support of the complex RB/A is Spec B. One
can prove that
f !M ∼
= RB/A ⊗LA M
for M ∈ Dbf (Mod A).

If the ring A is Gorenstein, then RA/K is a shift of an invertible A-module.
Hence:
Corollary 5.3. Assume that in Theorem 5.2 the ring A is Gorenstein. Then
RB/A is a dualizing complex over B
The rigid complex RB/A allows us to characterize Cohen-Macaulay homo-

morphisms, as follows.
Theorem 5.4. ([Ye2]) Let A be a tractable ring, and let f ∗ : A → B be an

essentially finite type flat homomorphism. Then the following conditions are
equivalent:
(i) f ∗ is an essentially Cohen-Macaulay homomorphism.

(ii) Let
&
Spec B = Spec Bi
i
be the decomposition into connected components. Then for any i there is

a finitely generated Bi -module ωBi /A , which is flat over A, and an integer
ni , such that

RB/A ∼
= ωBi /A [ni ]
i
in D(Mod B).
The module

ωB/A := ωBi /A
i
is called the relative dualizing module of f ∗ : A → B. Note that the complex

i ωBi /A [ni
] is rigid, but in general it is not a dualizing complex over B. Still
the fibers of i ωBi /A [ni ] are dualizing complexes – this can be seen by taking
A = k(p) in the next result, and using Corollary 5.3
Here is a “rigid” version of Conrad’s base change theorem [Co].
Theorem 5.5. ([Ye2]) Let

A −−−−→ B
⏐ ⏐
⏐ ⏐

A −−−−→ B
be a cartesian diagram of rings, i.e.
B ∼
= A ⊗A B,
with A and A tractable rings. Assume A → B is an essentially Cohen-

Macaulay homomorphism. (There isn’t any restriction on the homomorphism
A → A .) Then:
(1) A → B is an essentially Cohen-Macaulay homomorphism.

(2) There is a unique isomorphism of B -modules
ωB /A ∼
= A ⊗A ωB/A
which respects rigidity.

462 Amnon Yekutieli
From this we can easily deduce the next result.
Corollary 5.6. Let A be a tractable ring, and let f ∗ : A → B be an essentially

Cohen-Macaulay homomorphism. Then the following conditions are equiva-
lent:
(i) f ∗ is an essentially Gorenstein homomorphism.

(ii) ωB/A is an invertible B-module.
Remark 5.7. The recent paper [AI] contains results similar to Theorem 5.4
and Corollary 5.6, obtained by different methods, and without the requirement
that A is tractable.
References
[AK] A. Altman and S. Kleiman, “Introduction to Grothendieck Duality,” Lecture
Notes in Math. 20, Springer, 1970.
[AI] L.L. Avramov and S. Iyengar, Gorenstein algebras and Hochschild cohomol-
ogy, Michigan Math. J. 57 (2008), 17–35.
[AJL] L. Alonso, A. Jeremı́as and J. Lipman, Duality and flat base change on formal
schemes, in “Studies in Duality on Noetherian Formal Schemes and Non-
Noetherian Ordinary Schemes,” Contemp. Math. 244, Amer. Math. Soc., 1999,
3–90.
[Be] K. Behrend, Differential Graded Schemes I: Perfect Resolving Algebras, eprint
arXiv:math/0212225v1 [math.AG] at http://arXiv.org.
[Co] B. Conrad, “Grothendieck Duality and Base Change,” Lecture Notes in Math.
1750, Springer, 2000.
[Hi] V. Hinich, Homological algebra of homotopy algebras, Comm. Algebra 25
(1997), no. 10, 3291–3323.
[HS] R. Hübl and P. Sastry, Regular differential forms and relative duality, Amer. J.
Math. 115 (1993), no. 4, 749–787.
[HK] R. Hübl and E. Kunz, Regular differential forms and duality for projective
morphisms, J. Reine Angew. Math. 410 (1990), 84–108.
[Hu] R. Hübl, “Traces of Differential Forms and Hochschild Homology,” Lecture
Notes in Math. 1368, Springer, 1989.
[Ke] B. Keller, Deriving DG categories, Ann. Sci. École Norm. Sup. (4) 27 (1994),
no. 1, 63–102.
[Li] J. Lipman, “Residues and Traces of Differential Forms via Hochschild Homol-
ogy,” Contemporary Mathematics 61, Amer. Math. Soc., Providence, RI, 1987.
[Ne] A. Neeman, The Grothendieck duality theorem via Bousfield’s techniques and
Brown representability, J. Amer. Math. Soc. 9 (1996), no. 1, 205–236.
[RD] R. Hartshorne, “Residues and Duality,” Lecture Notes in Math. 20, Springer-
Verlag, Berlin, 1966.
[VdB] M. Van den Bergh, Existence theorems for dualizing complexes over non-
commutative graded and filtered ring, J. Algebra 195 (1997), no. 2, 662–679.
[Ye1] A. Yekutieli, “An Explicit Construction of the Grothendieck Residue Complex”
(with an appendix by P. Sastry), Astérisque 208 (1992).
[Ye2] A. Yekutieli, Rigidity, Residues, and Grothendieck Duality for Schemes, in
preparation.
[YZ1] A. Yekutieli and J.J. Zhang, Rings with Auslander dualizing complexes, J.
Algebra 213 (1999), no. 1, 1–51.
[YZ2] A. Yekutieli and J.J. Zhang, Rigid Dualizing Complexes and Perverse Sheaves
over Differential Algebras, Compositio Math. 141 (2005), 620–654.
[YZ3] A. Yekutieli and J.J. Zhang, Dualizing Complexes and Perverse Sheaves on
Noncommutative Ringed Schemes, Selecta Math. 12 (2006), 137–177.
[YZ4] A. Yekutieli and J.J. Zhang, Rigid Complexes via DG Algebras, Trans. AMS
360 no. 6 (2008), 3211–3248.
[YZ5] A. Yekutieli and J.J. Zhang, Rigid Dualizing Complexes over Commutative
Rings, Algebr. Represent. Theory, 12, no. 1 (2009), 19–52.
Department of Mathematics Ben Gurion University, Be’er Sheva 84105, Israel

E-mail address:

(Holm T., Jorgensen P., Rouquier R. (Eds.) ) Triang (BookFi)

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

(Holm T., Jorgensen P., Rouquier R. (Eds.) ) Triang (BookFi)

Uploaded by

Copyright:

Available Formats

This page intentionally left blank

LONDON MATHEMATICAL SOCIETY LECTURE NOTE SERIES

224 Computability, enumerability, unsolvability, S. B. COOPER, T. A. SLAMAN & S. S. WAINER (eds)

This publication is in copyright. Subject to statutory exception

First published 2010

Printed in the United Kingdom at the University Press, Cambridge

Library of Congress Cataloguing in Publication data

ISBN 978-0-521-74431-7 Paperback

Cambridge University Press has no responsibility for the persistence or

Preface page vii

Triangulated categories: definitions, properties, and examples 1

Derived categories of coherent sheaves on algebraic varieties 408

This volume grew out of a Workshop on Triangulated Categories held at the

Hannover, Newcastle, and Oxford, January 2010

Triangulated categories were introduced in the mid 1960’s by J.L. Verdier

Definition 1.1. A category A is called an additive category if the following

HomA (Y, Z) × HomA (X, Y ) → HomA (X, Z)

is bilinear over the integers.

ring elements f1 , f2 there existed a unique element f factoring them as

1.1. The category of complexes

The complexes over an additive category A together with the morphisms of

Proposition 1.4. Let A be an additive category. Then the category of complexes

From uniqueness in the universal property applied to

1.2. The homotopy category of complexes

In particular, setting g to be the zero morphism, we can speak of morphisms

(fn − gn )αn = (dn+1

Similarly, if f, g : X → Y are homotopic and β : Y → Z is a morphism of

Definition 1.6. Let A be an additive category. The homotopy category K(A)

HomK(A) (X, Y ) := HomC(A) (X, Y )/ ∼ .

Proposition 1.7. Let A be an additive category. Then the homotopy category

Proof. Addition of morphisms in K(A) is defined via addition on representa-

By the usual universal property argument, the kernel, if it exists, is unique up

Again, the cokernel, if it exists, is unique up to isomorphism; notation:

Note that k̃ ◦ g̃ ◦ k = f ◦ k = 0, which implies that g̃ ◦ k : ker f → im f must

and deduce that there is a unique morphism coim f → im f , as desired.

Definition 2.3. An additive category A is called an abelian category if the

(A4) Every morphism in A has a kernel and a cokernel.

Proposition 2.5. Let A be an abelian category. Then the category of complexes

Proof. We have seen in Proposition 1.4 that C(A) is an additive category, so it

kn−1 ◦ dnK ◦ dn+1

since X is a complex. By uniqueness of the map in the universal property of

The reverse implication is obvious.

An important observation is that the homotopy category K(A) is not abelian

3. Definition of triangulated categories

A morphism of triangles is a triple (f, g, h) of morphisms such that the

If in this situation, the morphisms f, g and h are isomorphisms in T , then the

Definition 3.1. A triangulated category is an additive category T together with

(TR0) Any triangle isomorphic to a distinguished triangle is again a distin-

(TR2) For every morphism f : X → Y in T there is a distinguished triangle

can be completed to a morphism of triangles (but not necessarily

The following version (TR5”) of the octahedral axiom can be found in

4. Some formal properties of triangulated categories

The left hand square is commutative by assumption on f . By (TR4) there exists

Proposition 4.3 (Triangulated 5-lemma). Suppose we are given a morphism

If f and g are isomorphisms then also h is an isomorphism.

Proof. We apply the functor Hom(Z , −) := HomT (Z , −) to the distinguished

Hom(Z , X) - Hom(Z , Y ) - Hom(Z , Z) - Hom(Z , X) - Hom(Z , Y )

By assumption, f and g are isomorphisms and hence also f∗ , g∗ , f∗ and

Remark 4.5. The notion of split monomorphism is synonymous with that of

By (TR3) and (TR4) it can be completed to a morphism of triangles, i.e. there

5. Abelian categories vs. triangulated categories

Definition 5.1. An abelian category A is called semisimple if every short exact

Theorem 5.3. Let T be a category which is triangulated and abelian. Then T

Hom(Z , X) - Hom(Z , Y ) - Hom(Z , Z) - Hom(Z , X) - Hom(Z , Y )

By assumption, f and g are isomorphisms and hence also f∗ , g∗ , f∗ and

Hence, the functor F, if it exists, is unique.